Image analysis method and system

TECHNICAL FIELD

The embodiments relate to the image processing field, and in particular, to an image analysis method and system.

BACKGROUND

A task of an image description is generating a corresponding text description for a given image. The image description may automatically extract information from the image and generate the corresponding text description based on the automatically extracted information, to translate the image into knowledge. For example, the picture description may generate a text description such as “a man is surfing on the sea” for an image shown in FIG. 1A.

Currently, image description can only perform a low-level semantic description on the image. Only a single-subject single action (for example, the man is surfing on the sea in FIG. 1A) or a multi-subject single action (for example, a group of students are doing morning exercises in FIG. 1B) can be described. However, image description cannot perform a panoramic semantic description on the image. Relationships between a plurality of subjects, relationships between the subjects and actions, and relationships between the actions (for example, a man sees that a woman is knocked down by a vehicle in FIG. 1C) cannot be described.

SUMMARY

The embodiments provide an image analysis method and system, to perform a panoramic semantic description on an image.

According to a first aspect, an image analysis method is provided, including:

obtaining influencing factors of t frames of images, where the influencing factors include self-owned features of h target subjects in each of the t frames of images and relational vector features between the h target subjects in each of the t frames of images, self-owned features of each target subject include a location feature, an attribute feature, and a posture feature, t and h are natural numbers greater than 1, the location feature is used to indicate a location of a corresponding target subject in the image, the attribute feature is used to indicate an attribute of the corresponding target subject, the posture feature is used to indicate an action of the corresponding target subject, and the relational vector features are used to indicate relationships between target subjects; and

obtaining a panoramic semantic description based on the influencing factors, where the panoramic semantic description includes a description of the relationships between target subjects, relationships between actions of the target subjects and the target subjects, and relationships between the actions of the target subjects.

In the solution, the higher-level panoramic semantic description can be obtained based on location features, attribute features, and posture features of a plurality of target subjects in a plurality of frames of images and relational vector features between the plurality of target subjects in the plurality of frames of images, to better reflect relationships between the plurality of subjects, relationships between the subjects and actions, and relationships between the actions in the images.

In some possible implementations, the obtaining influencing factors of a panoramic semantic description includes:

- extracting features of the t frames of images to obtain t feature vectors;
- extracting location features of the t feature vectors to obtain the location features;
- extracting attribute features of the t feature vectors to obtain the attribute features;
- extracting posture features of the t feature vectors to obtain the posture features; and
- extracting relational vector features of the t feature vectors to obtain the relational vector features.

In some possible implementations, the location features, the attribute features, the posture features, and the relational vector features are extracted by a same convolutional neural network (CNN).

In the solution, the same CNN is used to extract the location features, the attribute features, the posture features, and the relational vector features. Therefore, when the location features, the attribute features, the posture features, and the relational vector features are extracted, previously extracted feature vectors may be used, to avoid extracting the vector features for a plurality of times, and reduce a calculation amount. There is no need to extract the feature vectors when the location features are extracted, extract the feature vectors when the attribute features are extracted, extract the feature vectors when the posture features are extracted, or extract the feature vectors when the relational vector features are extracted.

In some possible implementations, region-of-interest pooling is performed on a feature vector i based on a target subject a and a target subject b that are in an image i to obtain a feature vector V_a,bcorresponding to the target subject a and the target subject b, where i, a, and b are all natural numbers, 0<i≤t, 1≤a,b≤h, and the feature vector i is extracted based on the image i.

Region-of-interest pooling is performed based on the target subject a to obtain a feature vector v_a,acorresponding to the target subject a.

A relational vector feature Vⁱ_abbetween the target subject a and the target subject b that are in the image i is calculated according to the following formula:

$G_{a, b} = \frac{1}{\sum  v_{a, b} } (w_{a, b} (v_{a, b}, v_{a, a}) v_{a, b})$

w_a,b=sigmoid(w(v_a,b, v_a,a)), sigmoid( ) is an S-type function, v_a,bis the feature vector corresponding to the target subject a and the target subject b, v_a,ais the feature vector corresponding to the target subject a, and w( ) is an inner product function.

In some possible implementations, the obtaining a panoramic semantic description based on the influencing factors includes:

extracting a first semantic description based on the location features;

extracting a second semantic description based on the attribute features and the first semantic description;

extracting a third semantic description based on the posture features and the second semantics; and

extracting the panoramic semantic description based on the relational vector features and the third semantic description.

In some possible implementations, the first semantic description, the second semantic description, and the third semantic description are extracted by a same recurrent neural network (RNN).

According to a second aspect, an image analysis system including a feature extraction module and a panoramic semantic model is provided.

The feature extraction module is configured to obtain influencing factors of a panoramic semantic description. The influencing factors include self-owned features of h target subjects in each of t frames of images and relational vector features between the h target subjects in each of the t frames of images, the self-owned features include a location feature, an attribute feature, and a posture feature, where t and h are natural numbers greater than 1. The location feature is used to indicate a location of a corresponding target subject in the image. The attribute feature is used to indicate an attribute of the corresponding target subject. The posture feature is used to indicate an action of the corresponding target subject. The relational vector features are used to indicate relationships between target subjects.

The panoramic semantic model is configured to obtain the panoramic semantic description based on the influencing factors. The panoramic semantic description includes a description of the relationships between target subjects, relationships between the target subjects and actions, and relationships between the actions.

In some possible implementations, the feature extraction module includes a feature vector extraction unit, a location feature extraction unit, an attribute feature extraction unit, a posture feature extraction unit, and a relational vector feature unit.

The feature vector extraction unit is configured to extract features of the t frames of images to obtain t feature vectors.

The location feature extraction unit is configured to extract location features of the t feature vectors to obtain the location features.

The attribute feature extraction unit is configured to extract attribute features of the t feature vectors to obtain the attribute features.

The posture feature extraction unit is configured to extract posture features of the t feature vectors to obtain the posture features.

The relational vector feature unit is configured to extract relational vector features of the t feature vectors to obtain the relational vector features.

In some possible implementations, the feature extraction module includes a CNN. The feature vector extraction unit, the location feature extraction unit, the attribute feature extraction unit, the posture feature extraction unit, and the relational vector feature extraction unit are integrated into the CNN.

In some possible implementations, the relational vector feature extraction unit is configured to: perform region-of-interest pooling on a feature vector i based on a target subject a and a target subject b that are in an image i to obtain a feature vector v_a,bcorresponding to the target subject a and the target subject b, where i, a, and b are natural numbers, 0<i≤t, and 1≤a,b≤h;

perform region-of-interest pooling based on the target subject a to obtain a feature vector v_a,acorresponding to the target subject a; and

calculate a relational vector feature Vⁱ_abbetween the target subject a and the target subject b that are in the image i according to the following formula:

$G_{a, b} = \frac{1}{\sum  v_{a, b} } (w_{a, b} (v_{a, b}, v_{a, a}) v_{a, b})$

w_a,b=sigmoid(w(v_a,b,v_a,a)), sigmoid( ) is an S-type function, v_a,bis the feature vector corresponding to the target subject a and the target subject b, v_a,ais the feature vector corresponding to the target subject a, and w( ) is an inner product function.

In some possible implementations, the panoramic semantic model includes a first time sequence feature extraction unit, a second time sequence feature extraction unit, a third time sequence feature extraction unit, and a fourth time sequence feature extraction unit.

The first time sequence feature extraction unit is configured to extract a first semantic description based on the location features.

The second time sequence feature extraction unit is configured to extract a second semantic description based on the attribute features and the first semantic description.

The third time sequence feature extraction unit is configured to extract a third semantic description based on the posture features and the second semantics.

The fourth time sequence feature extraction unit is configured to extract the panoramic semantic description based on the relational vector features and the third semantic description.

In some possible implementations, the panoramic semantic model includes a RNN. The first time sequence feature extraction unit, the second time sequence feature extraction unit, the third time sequence feature extraction unit, and the fourth time sequence feature extraction unit are respectively one layer in the RNN.

According to a third aspect, a compute node including a processor and a memory is provided. The processor is configured to:

obtain influencing factors of t frames of images, where the influencing factors include self-owned features of h target subjects in each of the t frames of images and relational vector features between the h target subjects in each of the t frames of images, self-owned features of each target subject include a location feature, an attribute feature, and a posture feature, t and h are natural numbers greater than 1, the location feature is used to indicate a location of a corresponding target subject in the image, the attribute feature is used to indicate an attribute of the corresponding target subject, the posture feature is used to indicate an action of the corresponding target subject, and the relational vector features are used to indicate relationships between target subjects; and

obtain a panoramic semantic description based on the influencing factors, where the panoramic semantic description includes a description of the relationships between target subjects, relationships between actions of the target subjects and the target subjects, and relationships between the actions of the target subjects.

In some possible designs, the processor is configured to:

- extract features of the t frames of images to obtain t feature vectors;
- extract location features of the t feature vectors to obtain the location features;
- extract attribute features of the t feature vectors to obtain the attribute features;
- extract posture features of the t feature vectors to obtain the posture features; and
- extract relational vector features of the t feature vectors to obtain the relational vector features.

In some possible implementations, the location features, the attribute features, the posture features, and the relational vector features are extracted by a same CNN.

In the solution, the same CNN is used to extract the location features, the attribute features, the posture features, and the relational vector features. Therefore, when the location features, the attribute features, the posture features, and the relational vector features are extracted, previously extracted feature vectors may be used, to avoid extracting the vector features for a plurality of times, and reduce a calculation amount. In other words, there is no need to extract the feature vectors when the location features are extracted, extract the feature vectors when the attribute features are extracted, extract the feature vectors when the posture features are extracted, or extract the feature vectors when the relational vector features are extracted.

In some possible implementations, region-of-interest pooling is performed on a feature vector i based on a target subject a and a target subject b that are in an image i to obtain a feature vector v_a,bcorresponding to the target subject a and the target subject b, where i, a, and b are all natural numbers, 0<i≤t, 1≤a,b≤h, and the feature vector i is extracted based on the image i.

Region-of-interest pooling is performed based on the target subject a to obtain a feature vector v_a,acorresponding to the target subject a.

A relational vector feature Vⁱ_abbetween the target subject a and the target subject b that are in the image i is calculated according to the following formula:

$G_{a, b} = \frac{1}{\sum  v_{a, b} } (w_{a, b} (v_{a, b}, v_{a, a}) v_{a, b})$

w_a,b=sigmoid(w(v_a,b,v_a,a)), sigmoid( ) is an S-type function, v_a,bis the feature vector corresponding to the target subject a and the target subject b, v_a,ais the feature vector corresponding to the target subject a, and w( ) is an inner product function.

In some possible implementations, the processor is configured to:

extract a first semantic description based on the location features; extract a second semantic description based on the attribute features and the first semantic description;

extract a third semantic description based on the posture features and the second semantics; and

extract the panoramic semantic description based on the relational vector features and the third semantic description.

In some possible implementations, the first semantic description, the second semantic description, and the third semantic description are extracted by a same RNN.

According to a fourth aspect, a compute node cluster including at least one compute node is provided. Each compute node includes a processor and a memory. The processor executes code in the memory to perform the method according to any one of the first aspect.

According to a fifth aspect, a computer program product is provided. When the computer program product is read and executed by a computer, the method according to any one of the first aspect is performed.

According to a sixth aspect, a computer non-transient storage medium including an instruction is provided. When the instruction is run on at least one compute node in a compute node cluster, the compute node cluster is enabled to perform the method according to any one of the first aspect.

BRIEF DESCRIPTION OF DRAWINGS

To describe the solutions in embodiments or in the background more clearly, the following describes the accompanying drawings for describing the embodiments or the background.

FIG. 1A is a schematic diagram of an image used for image descriptions;

FIG. 1B is a schematic diagram of an image used for image descriptions;

FIG. 1C is a schematic diagram of an image used for image descriptions;

FIG. 2 is a schematic diagram of a single-frame image used for a panoramic semantic description according to an embodiment;

FIG. 3 is a schematic diagram of a plurality of frames of images used for a panoramic semantic description according to an embodiment;

FIG. 4 is a schematic diagram of feature extraction of location features, attribute features, posture features, and relational vector features;

FIG. 5 is a schematic diagram of a panoramic semantic model according to an embodiment;

FIG. 6A is a schematic diagram of a panoramic semantic model according to another embodiment;

FIG. 6B is a schematic diagram of a panoramic semantic model according to another embodiment;

FIG. 7 is a flowchart of a semantic description method according to an embodiment;

FIG. 8 is a schematic diagram of a structure of a semantic description system according to an embodiment;

FIG. 9 is a schematic diagram of a structure of a compute node according to an embodiment;

FIG. 10 is a schematic diagram of a structure of a cloud service cluster according to an embodiment;

FIG. 11 is a schematic diagram of a structure of a semantic description system according to another embodiment; and

FIG. 12 is a schematic diagram of a structure of a semantic description system according to still another embodiment.

DETAILED DESCRIPTION OF THE EMBODIMENTS

Terms used in embodiments are merely used to explain the embodiments but are non-limiting.

A single image used for a panoramic semantic description in the embodiments is first described in detail.

FIG. 2 is a schematic diagram of a single-frame image used for a panoramic semantic description according to an embodiment. In the embodiment, the single-frame image used for the panoramic semantic description may include a plurality of target subjects. The target subjects may be one or more of a person, an animal, an object, and the like. FIG. 2 is used as an example. The target subjects in the image shown in FIG. 2 include a man, a woman, and a vehicle. Different target subjects may perform different actions. The actions may be one or more of drinking water, reading a book, doing an exercise, playing basketball, playing football, running, swimming, and the like. FIG. 2 is used as an example. In the figure, an action of the man is looking toward the woman, an action of the woman is falling down, and an action of the vehicle in the figure is knocking the woman down. It may be understood that FIG. 2 is merely an example. In the embodiments, the target subject may be another subject, there may be more target subjects, and the action of the target subject may be another action or the like and are non-limiting.

In an embodiment, as shown in FIG. 3, an image analysis system may extract, from a video in a time sequence, t frames of images I₁, I₂, . . . , and I_tused for a panoramic semantic description, where t is a natural number. The images I₁, I₂, . . . , and I_tall include same target subjects. For example, the image I₁includes a target subject 1, a target subject 2, and a target subject 3. The image I₂includes the target subject 1, the target subject 2, and the target subject 3, . . . , and the image I_talso includes the target subject 1, the target subject 2, and the target subject 3. It may be understood that time intervals between two adjacent frames of images in the t frames of images may be same or different and is non-limiting.

In an embodiment, the image analysis system may perform the panoramic semantic description on the image I_tby using a panoramic semantic model. Input variables of the panoramic semantic model are influencing factors of the panoramic semantic description. The influencing factors of the panoramic semantic description include self-owned features (including location features, attribute features, and posture features) of target subjects in the images I₁to I_tand relational vector features between the target subjects in the images I₁to I_t.

The location feature may be used to indicate a location of a corresponding target subject in the corresponding image. The location feature may be represented as (x, y, w, h), where x and y are a horizontal coordinate and a vertical coordinate of a central point of the target subject in the image, w is a width of the target subject in the image, and h is a height of the target subject in the image. The attribute feature is used to indicate an attribute of the corresponding target subject. The attribute feature may include a plurality of types of attribute features. Attribute features may be different if target subjects are different. For example, the target subject is a person. The attribute feature of the target subject may include one or more of a gender, a hairstyle, a clothes type, a clothes color, a height, a body size, and the like. The posture feature of the target subject is used to indicate an action of the corresponding target subject. The target subject also has a plurality of posture features. Posture features may be different if target subjects are different. For example, the target subject is a person. The posture feature of the target subject may include one or more of falling down, lying down, walking, running, jumping, and the like. The relational feature vector is a vector representing a relationship between two target subjects.

For example, each frame of image in the images I₁, I₂, . . . , and I_tincludes h target subjects. The influencing factors of the panoramic semantic description may include:

influencing factors, of a panoramic semantic description, obtained from the image I₁: self-owned features of h target subjects in the image I₁and relational vector features between the h target subjects in the image I₁.

The self-owned features of the h target subjects in the image I₁include:

$[\begin{matrix} P_{1}^{1} & S_{1}^{1} & Z_{1}^{1} \\ P_{2}^{1} & S_{2}^{1} & Z_{2}^{1} \\ \dots & \dots & \dots \\ P_{h}^{1} & S_{h}^{1} & Z_{h}^{1} \end{matrix}]$

Herein, a location feature P₁¹, an attribute feature S₁¹, and a posture feature Z₁¹are self-owned features of a target subject 1 in the image I₁, a location feature P₂¹, an attribute feature S₂¹, and a posture feature Z₂¹are self-owned features of a target subject 2 in the image I₁, . . . , and a location feature P_h¹, an attribute feature S_h¹, and a posture feature Z_h¹are self-owned features of a target subject h in the image I₁.

The relational vector features between the h target subjects in the image I₁include:

$[\begin{matrix} V_{12}^{1} & V_{13}^{1} & \dots & V_{1 h}^{1} \\ V_{23}^{1} & \dots & V_{1 h}^{1} \\ ⋮ & ⋮ \\ V_{h - 1 h}^{1} \end{matrix}]$

Herein, V₁₂¹is a relational vector feature between the target subject 1 and the target subject 2 in the image I₁, V₁₃¹is a relational vector feature between the target subject 1 and a target subject 3 in the image I₁, . . . , V_1h¹is a relational vector feature between the target subject 1 and the target subject h in the image I₁, V₂₃¹is a relational vector feature between the target subject 2 and the target subject 3 in the image I₁, . . . , V_2h¹is a relational vector feature between the target subject 2 and the target subject h in the image I₁, . . . , and V_h-1h¹is a relational vector feature between a target subject (h-1) and the target subject h in the image I₁.

The influencing factors of the panoramic semantic description may further include: influencing factors of, a panoramic semantic description, obtained from the image I₂: self-owned features of h target subjects in the image I₂and relational vector features between the h target subjects in the image I₂.

The self-owned features of the h target subjects in the image I₂include:

$[\begin{matrix} P_{1}^{2} & S_{1}^{2} & Z_{1}^{2} \\ P_{2}^{2} & S_{2}^{2} & Z_{2}^{2} \\ \dots & \dots & \dots \\ P_{h}^{2} & S_{h}^{2} & Z_{h}^{2} \end{matrix}]$

Herein, a location feature P₁², an attribute feature S₁², and a posture feature Z₁²are self-owned features of a target subject 1 in the image I₂, a location feature P₂², an attribute feature S₂², and a posture feature Z₂²are self-owned features of a target subject 2 in the image I₂, . . . , and a location feature P_h², an attribute feature S_h², and a posture feature Z_h²are self-owned features of a target subject h in the image I₂.

The relational vector features between the h target subjects in the image I₂include:

$[\begin{matrix} V_{12}^{2} & V_{13}^{2} & \dots & V_{1 h}^{2} \\ V_{23}^{2} & \dots & V_{1 h}^{2} \\ ⋮ & ⋮ \\ V_{h - 1 h}^{2} \end{matrix}]$

Herein, V₁₂²is a relational vector feature between the target subject 1 and the target subject 2 in the image I₂, V₁₃²is a relational vector feature between the target subject 1 and a target subject 3 in the image I₂, . . . , V_1h²is a relational vector feature between the target subject 1 and the target subject h in the image I₂, V₂₃²is a relational vector feature between the target subject 2 and the target subject 3 in the image I₂, V_2h²is a relational vector feature between the target subject 2 and the target subject h in the image I₂, . . . , and V_h-1h²is a relational vector feature between a target subject (h-1) and the target subject h in the image I₂.

The influencing factors of the panoramic semantic description may further include: influencing factors, of a panoramic semantic description, obtained from the image I_t: self-owned features of h target subjects in the image I_tand relational vector features between the h target subjects in the image I_t.

The self-owned features of the h target subjects in the image I_tinclude:

$[\begin{matrix} P_{1}^{t} & S_{1}^{t} & Z_{1}^{t} \\ P_{2}^{t} & S_{2}^{t} & Z_{2}^{t} \\ \dots & \dots & \dots \\ P_{h}^{t} & S_{h}^{t} & Z_{h}^{t} \end{matrix}]$

Herein, a location feature P₁^t, an attribute feature S₁^t, and a posture feature Z₁^tare self-owned features of a target subject 1 in the image I_t, a location feature P₂^t, an attribute feature S₂^t, and a posture feature Z₂^tare self-owned features of a target subject 2 in the image I_t, . . . , and a location feature P_h^t, an attribute feature S_h^t, and a posture feature Z_h^tare self-owned features of a target subject h in the image I_t.

The relational vector features between the h target subjects in the image I_tinclude:

$[\begin{matrix} V_{12}^{t} & V_{13}^{t} & \dots & V_{1 h}^{t} \\ V_{23}^{t} & \dots & V_{2 h}^{t} \\ ⋮ & ⋮ \\ V_{h - 1 h}^{t} \end{matrix}]$

Herein, V^t₁₂is a relational vector feature between the target subject 1 and the target subject 2 in the image I_t, V₁₃^tis a relational vector feature between the target subject 1 and a target subject 3 in the image I_t, . . . , V_1h^tis a relational vector feature between the target subject 1 and the target subject h in the image I_t, V₂₃^tis a relational vector feature between the target subject 2 and the target subject 3 in the image I_t, . . . , V_2h^tis a relational vector feature between the target subject 2 and the target subject h in the image I_t, . . . , and V_h-1h^tis a relational vector feature between a target subject (h−1) and the target subject h in the image I_t.

It may be understood that the example of the influencing factors of the panoramic semantic description is merely used as an example. The influencing factors of the panoramic semantic description may further include other influencing factors and is non-limiting.

In an exemplary embodiment, location features, attribute features, and posture features of the target subjects in the images I₁, I₂, . . . , and I_tand relational vector features between the target subjects in the images I₁, I₂, . . . , and I_tmay be separately calculated based on feature vectors V₁, V₂, . . . , and V_tof the images I₁, I₂, . . . , and V_t. In other words, location features, attribute features, and posture features of target subjects in the image I₁and relational vector features between the target subjects in the image I₁may be calculated based on the feature vector V₁of the image I₁, location features, attribute features, and posture features of target subjects in the image I₂and relational vector features between the target subjects in the image I₂may be calculated based on the feature vector V₂of the image I₂, . . . , and location features, attribute features, and posture features of target subjects in the image I_tand relational vector features between the target subjects in the image I_tmay be calculated based on the feature vector V_tof the image I_t.

As shown in FIG. 4, feature vectors V₁, V₂, . . . , and V_tof images I₁, I₂, . . . , and I_tmay be obtained in this way. An image I_iis used as an example. The image I_imay be input into a feature vector extraction unit to obtain a feature vector V_iof the image I_i, where i is a natural number, and 1≤i≤t. The feature vector extraction unit may sequentially include an input layer, a convolution compute layer, a pooling layer, and a fully connected layer.

The input layer:

It is assumed that input of the input layer is the image I_i, and output is equal to the input. In other words, no processing is performed on the input. For ease of description, it is assumed herein that there is no processing in the input layer. However, processing such as normalization may be performed in the input layer.

The convolution compute layer:

The image I_ioutput by the input layer is used as input of the convolution compute layer, and is convolved by n convolution kernels K_l(l=1, 2, . . . , n) to generate n feature images a_l(l=1, 2, . . . , n). A process of generating each feature image a_lis specifically as follows:

$C_{l} = conv 2 (I, K_{l}, {}^{'}{valid}^{'},) + b_{l}$

$u_{l} = C_{l}$

$a_{l} = f (u_{l})$

conv represents performing a convolution operation on the image I by using a convolution kernel K_l, valid represents a padding manner, b_lrepresents an offset value, u_lrepresents a convolution calculation result, and f( ) represents an activation function such as a relu function.

The pooling layer:

The n feature images a_loutput by the convolution compute layer are used as input of the pooling layer, and after pooling is performed by using a pooling window, n pooled images b_l(l=1, 2, . . . , n) are generated. A generation process of each pooled image b_lis specifically as follows:

b_l=max Pool (a_l)

maxPool represents average pooling.

The fully connected layer:

The n pooled images b_l(l=1, 2, . . . , n) are sequentially expanded into a vector, and are sequentially connected into a long vector. The long vector is used as input of the fully connected layer network. Output of the fully connected layer is the feature vector V_iof the image I_i.

The parameters in the feature vector extraction unit, the convolution kernel K_l(including an element, a size, a step size, and the like), the offset value b_l, f( ) and β_lmay be manually set based on a feature (a location feature, an attribute feature, a posture feature, and a relational vector feature) that needs to be extracted, a size of the image I_i, and the like. The convolution kernel K_lis used as an example. When the feature that needs to be extracted is the location feature, the element of the convolution kernel K_lmay be an element of a sobel operator. For another example, when the image I_iis relatively large, the size of the convolution kernel K_lmay be relatively large. On the contrary, when the image I_iis relatively small, the size of the convolution kernel K_lmay be relatively small. For another example, when the image I_iis relatively large, the step size of the convolution kernel K_lmay also be relatively large. On the contrary, when the image I_iis relatively small, the step size of the convolution kernel K_lmay also be relatively small.

It may be understood that the feature vector extraction unit is merely used as an example. The feature vector extraction unit may alternatively be in another form. For example, the feature vector extraction unit may include more convolution compute layers and more pooling layers, and may fill the image I_i, and the like.

For brevity, the foregoing describes only extraction of the feature vector V_iof the image I_i. Actually, extraction manners of the feature vectors V₁, V₂, . . . , and V_lof the images I₁, I₂, . . . , and I_tare similar to the extraction manner of the feature vector V_iof the image I_i. Details are not described herein again.

As shown in FIG. 4, location features of target subjects in the images I₁, I₂, . . . , and I_tmay be obtained in this way. The image I_iis used as an example. It may be assumed that the image I_iincludes the h target subjects. The feature vector V_iis input into a location feature extraction unit for extraction to obtain location features P₁ⁱ, P₂ⁱ, . . . , and P_hⁱof the h target subjects in the image I_i, where i is a natural number, and 1≤i≤t. The location feature extraction unit may be represented as:

y₁=g₁(x₁).

Herein, x₁may be the feature vector V_iof the image I_i, y₁may be the location features P₁ⁱ, P₂ⁱ, . . . , and P_hⁱof the h target subjects in the image I_i, g₁( ) is a mapping relationship between the feature vector V_iand the location features P₁ⁱ, P₂ⁱ, . . . , and P_hⁱ, and g₁( ) may be obtained by training a large quantity of known images and location features of known target subjects. For brevity, the foregoing describes only extraction of the location features P₁ⁱ, P₂ⁱ, . . . and P_hⁱof the h target subjects in the image I_i. Actually, extraction manners of location features P₁¹, P₂¹, . . . , and P_h¹, P₁², P₂², . . . , and P_h², . . . , and P₁^t, P₂^t, . . . , and P_h^tof h target subjects in each of the images I₁, I₂, . . . , and I_tare similar to the extraction manner of the location features P₁ⁱ, P₂ⁱ, . . . , and P_hⁱof the h target subjects in the image I_i. Details are not described herein again.

As shown in FIG. 4, attribute features of the target subjects in the images I₁, I₂, . . . , and I_tmay be obtained in this way. The image I_iis used as an example. It is assumed that the image I_iincludes the h target subjects. The feature vector V_iis input into an attribute feature extraction unit for extraction to obtain attribute features S₁ⁱ, S₂ⁱ, . . . , and S_hⁱof the h target subjects in the image I_i, where i is a natural number, and 1≤i≤t. The attribute extraction unit may be represented as:

y₂=g₂(x₁)

Herein, x₁may be the feature vector V_iof the image I_i, y₂may be the attribute features S₁ⁱ, S₂ⁱ, . . . , and S_hⁱof the h target subjects in the image I_i, g₂( ) is a mapping relationship between the feature vector V_iand the attribute features S₁ⁱ, S₂ⁱ, . . . , and S_hⁱ, and g₂( ) may be obtained by training a large quantity of known images and attribute features of known target subjects. For brevity, the foregoing describes only extraction of the attribute features S₁ⁱ, S₂ⁱ, . . . , and S_hⁱof the h target subjects in the image I_i. Actually, extraction manners of attribute features S₁¹, S₂¹, . . . , and S_h¹, S₁², S₂², . . . , and S_h², . . . , and S₁^t, S₂^t, . . . , and s_h^tof h target subjects in each of the images I₁, I₂, . . . , and I_tare similar to the extraction manner of the attribute features S₁ⁱ, S₂ⁱ, . . . , and S_hⁱof the h target subjects in the image I_i. Details are not described herein again.

As shown in FIG. 4, posture features of the target subjects in the images I₁, I₂. . . , and I_tmay be obtained in this way. The image I_iis used as an example. It is assumed that the image I_iincludes the h target subjects. The feature vector V_iis input into a posture extraction unit for extraction to obtain posture features Z₁ⁱ, Z₂ⁱ, . . . , and Z_hⁱof the h target subjects in the image I_i, where i is a natural number, and 1≤i≤t. The posture extraction unit may be represented as:

y₃=g₃(x₁).

Herein, x₁may be the feature vector V_iof the image I_i, y₃may be the posture features Z₁ⁱ, Z₂ⁱ, . . . , and Z_hⁱof the h target subjects in the image I_i, g₃( ) is a mapping relationship between the feature vector V_iand the posture features Z₁ⁱ, Z₂ⁱ, . . . , and 4, and g₂( ) may be obtained by training a large quantity of known images and posture features of known target subjects. For brevity, the foregoing describes only extraction of the posture features Z₁ⁱ, Z₂ⁱ, . . . , and Z_hⁱof the h target subjects in the image I_i. Actually, extraction manners of posture features Z₁¹, Z₂¹, . . . , and Z_h¹, Z₁², Z₂², . . . , and Z_h², . . . , and Z₁^t, Z₂^t, . . . , and Z_h^tof h target subjects in each of the images I₁, I₂, . . . , and I_tare similar to the extraction manner of the posture features Z₁ⁱ, Z₂ⁱ, . . . , and Z_hⁱof the h target subjects in the image I_i. Details are not described herein again.

As shown in FIG. 4, relational vector feature between the target subjects in the images I₁, I₂, . . . , and I_tmay be obtained in this way. The image I_iis used as an example. It is assumed that the image I_iincludes the h target subjects. Relational vector features between the h target subjects in the image I_iinclude: G₁₂ⁱ, G₁₃ⁱ, . . . , and G_1hⁱ; G₂₃ⁱ, . . . , and G_2hⁱ; . . . ; and G_h-1hⁱ. A relational feature vector G_abⁱmay be calculated by a relational vector feature extraction unit, where i, a, and b are natural numbers, 1≤i≤t and 1≤a,b≤h.

The relational vector feature extraction unit is configured to perform region-of-interest pooling (ROI) based on a target subject a and a target subject b to obtain a feature vector v_a,b, corresponding to the target subject a and the target subject b.

The relational vector feature extraction unit is configured to perform ROI pooling based on the target subject a to obtain a feature vector v_a,acorresponding to the target subject a.

The relational vector feature extraction unit is configured to calculate the relational vector feature V_abⁱaccording to the following formula:

$G_{a, b} = \frac{1}{\sum  v_{a, b} } (w_{a, b} (v_{a, b}, v_{a, a}) v_{a, b})$

For brevity, the foregoing describes only extraction of the relational vector features G₁₂ⁱ, G₁₃ⁱ, . . . , and G_h-1hⁱof the h target subjects in the image I_i. Extraction manners of relational vector features G₁₂¹, G₁₃¹, . . . , G_h-1h¹, G₁₂², G₁₃², . . . , and G_h-1h², . . . , and G₁₂^t, G₁₃^t, . . . , and G_h-1h^tof h target subjects in each of the images I₁, I₂, . . . , and I₁are similar to the extraction manner of the relational vector features G₁₂ⁱ, G₁₃ⁱ, . . . , and G_h-1hⁱof the h target subjects in the image I_i. Details are not described herein again.

Extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features may be implemented separately by different CNN, or may be integrated into a same CNN for implementation and is non-limiting. The CNN may include a VGGNet, a ResNet, an FPNet, and the like. When the extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features is integrated into the same CNN for completion, the extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features may be implemented separately at different layers in the CNN.

In an exemplary embodiment, influencing factors of a panoramic semantic description (the location features of the target subjects in the images I₁, I₂, . . . , and I_t, the attribute features of the target subjects in the images I₁, I₂, . . . , and I_t, the posture features of the target subjects in the images I₁, I₂, . . . , and I_t, and the relational vector features between the target subjects in the images I₁, I₂, . . . , and I_t) have the following influences on the panoramic semantic description: The location features of the target subjects in the images I₁, I₂, . . . , and I_tmay provide a first semantic description about locations of the target subjects. The attribute features of the target subjects in the images I₁, I₂, . . . , and I_tand the first semantic description may be used to obtain a second semantic description of attributes of the target subjects. The posture features of the target subjects in the images I₁, I₂, . . . , and I_tand the second semantic description may be used to obtain a third semantic description. Finally, the relational vector features between the target subjects of the images I₁, I₂, . . . , and I_tand the third semantic description may be used to obtain the panoramic semantic description.

An example shown in FIG. 3 is used as an example. The influencing factors of the panoramic semantic description have the following influences on the panoramic semantic description: First, a first semantic description “an object A and an object B are on the left side of an object C” may be obtained by using location features of each of a man, a woman, and a vehicle that are in the images I₁, I₂, . . . , and I_tin FIG. 3. Then, a second semantic description “the woman and the vehicle are on the left side of the man” may be obtained by using attribute features of the man, the woman, and the vehicle that are in the images I₁, I₂, . . . , and I_tin FIG. 3 and the first semantic description. Then, a third semantic description may be obtained by using posture features of the man, the woman, and the vehicle that are in the images I₁, I₂, . . . , and I_tin FIG. 3 and the second semantic description. Finally, the panoramic semantic description “the man on the right sees that the woman on the left is knocked down by the vehicle” may be obtained by using relational vector features in the images I₁, I₂, . . . , and I_tin FIG. 3 and the third semantic description.

It may be understood that the example shown in FIG. 3 is merely used as an example. In another embodiment, the panoramic semantic description may be further performed on another image.

In an embodiment, a panoramic semantic model may be represented as:

y=Panorama(x).

x is the influencing factors of the panoramic semantic description, y is the panoramic semantic description, Panorama( ) is a mapping relationship between the influencing factors of the panoramic semantic description and the panoramic semantic description, and Panorama( ) may be obtained by training a large quantity of influencing factors of known panoramic semantic descriptions and the known panoramic semantic descriptions. In an exemplary embodiment, a panoramic semantic model may be shown in FIG. 5.

Location features P₁¹, P₂¹, . . . , and P_h¹, P₁², P₂², . . . , and P_h², . . . , and P₁^t, P₂^t, . . . , and P_h^tof h target subjects in images I₁, I₂, . . . , and I_tare input into a time sequence feature extraction unit 1 to obtain a first semantic description.

Attribute features S₁¹, S₂¹, . . . , and S_h¹, S₁², S₂², . . . , and S_h², . . . , and S₁^t, S₂^t, . . . , and S_h^tof the h target subjects in the images I₁, I₂, . . . , and I_tand the first semantic description are input into a time sequence feature extraction unit 2 to obtain a second semantic description.

Posture features Z₁¹, Z₂¹, . . . , and Z_h¹, Z₁², Z₂², . . . , and Z_h², . . . , and Z₁^t, Z₂^t, . . . , and Z_h^tof the h target subjects in the images I₁, I₂, . . . , and I_tand the second semantic description are input into a time sequence feature extraction unit 3 to obtain a third semantic description.

Relational vector features V₁₂¹, V₁₃¹, . . . , and V_h-1h¹, V₁₂², V₁₃², . . . , and V_h-1h², . . . , and V₁^t, V₂^t, . . . , and V_h^tof the h target subjects in the images I₁, I₂, . . . , and I_tand the third semantic description are input into a time sequence feature extraction unit 4 to obtain a panoramic semantic description.

It may be understood that extraction of the first semantic description, the second semantic description, the third semantic description, and the panoramic semantic description may be implemented separately by different RNN, or may be implemented by a same RNN. The RNN may include a long short-term memory (LSTM) model, a bidirectional LSTM (BiLSTM) model, and the like. When the extraction of the first semantic description, the second semantic description, the third semantic description, and the panoramic semantic description is implemented in the same RNN, the extraction may be implemented separately at different layers in the RNN.

For example, time sequence feature extraction units 1 to 4 are all LSTMs. In a embodiment, a panoramic semantic model may be shown in FIG. 6A and FIG. 6B.

The following provides calculation processes in a neuron 1, a neuron 2, . . . , and a neuron tin the first time sequence feature extraction unit.

In the neuron 1:

First, a forgotten value f₁₀is calculated based on location features P₁¹, P₂¹, . . . , and P_h¹of h target subjects in an image I₁and an initial output value h₁₀:

f₁₀=σ(W₁₀^f·[h₁₀,[P₁¹,P₂¹, . . . P_h¹]]+b₁₀^f).

σ( ) is a sigmoid function, b₁₀^fis an offset value, and W₁₀^fis a weight matrix.

Then, an input value C₁₁is calculated based on the location features P₁¹, P₂¹, . . . , and P_h¹of the h target subjects in the image I₁, an initial input value C₁₀, the initial output value h₁₀, and the forgotten value f₁₀:

i₁₀=σ(W₁₀ⁱ·[h₁₀,[P₁¹,P₂¹, . . . P_h¹]]+b₁₀ⁱ)
{tilde over (C)}₁₀=tanh(W₁₀^c·[h₁₀,[P₁¹,P₂¹, . . . P_h¹]]+b₁₀^c)
C₁₁=f₁₀*C₁₀+i₁₀*{tilde over (C)}₁₀

σ( ) is a sigmoid function, tanh is an activation function, both W₁₀ⁱand W₁₀^care weight matrices, and both b₁₀ⁱand b₁₀^care offset values.

Finally, an output value h₁₁is calculated based on the location features P₁¹, P₂¹, . . . , and P_h¹of the h target subjects in the image I₁, the initial output value h₁₀and the input value C₁₁:

o₁₀=σ(W₁₀^o·[h₁₀,[P₁¹,P₂¹, . . . P_h¹]]+b₁₀^o)
h₁₁=o₁₀tanh(C₁₁)

σ( ) is a sigmoid function, tanh is an activation function, W₁₀^ois a weight matrix, and b₁₀^ois an offset value.

In the neuron 2:

First, a forgotten value f₁₁is calculated based on location features P₁², P₂², . . . , and P_h²of h target subjects in an image I₂and the output value h₁₁;

f₁₁=σ(W₁₁^f·[h₁₁,[P₁²,P₂², . . . P_h²]]+b₁₁^f)

σ( ) is a sigmoid function, b₁₁^fis an offset value, and W₁₁^fa weight matrix.

Then, an input value C₁₂is calculated based on the location features P₁², P₂², . . . , and P_h²of the h target subjects in the image I₂an input value C₁₁, the output value h₁₁, and the forgotten value f:

i₁₁=σ(W₁₁ⁱ·[h₁₁,[P₁²,P₂², . . . P_h¹]]+b₁₁ⁱ)
{tilde over (C)}₁₁=tanh(W₁₁^c·[h₁₁,[P₁²,P₂², . . . P_h²]]+b₁^c)
C₁₂=f₁₁*C₁₁+i₁₁*{tilde over (C)}₁₁

σ( ) is a sigmoid function, tanh is an activation function, both W₁₁ⁱand W₁₁^care weight matrices, and both b₁₁ⁱand b₁₁^care offset values.

Finally, an output value h₁₂is calculated based on the location features P₁², P₂², . . . , and P_h²of the h target subjects in the image I₂, the output value k₁and the input value C₁₂:

o₁₁=σ(W₁₁^o·[h₁₁,[P₁²,P₂², . . . P_h²]]+b₁₁^o)
h₁₂=o₁₁tanh(C₁₂)

σ( ) is a sigmoid function, tanh is an activation function, W₁₁^ois a weight matrix, and b₁₁^ois an offset value.

In the neuron t:

First, a forgotten value f_1t-1is calculated based on location features P₁^t, P₂^t, . . . , and P_h^tof h target subjects in an image I_tand an output value h_1t-1:

f_1t-1=σ(W_1t-1^f·[h_1t-1,[P₁^t,P₂^t, . . . P_h^t]]+b_1t-1^f)

σ( ) is a sigmoid function, b_1t-1^fis an offset value, and W_1t-1^fis a weight matrix.

Then, an input value C_1tis calculated based on the location features P₁^t, P₂^t, . . . , and P_h^tof the h target subjects in the image I_t, an input value C_1t-i, the output value h_1t-1, and the forgotten value f_1t-1:

i_1t-1=σ(W_1t-1ⁱ·[h_1t-1,[P₁^t,P₂^t, . . . P_h^t]]+b_1t-1ⁱ)
{tilde over (C)}_1t-1=tanh(W_1t-1^c·[h_1t-1,[P₁^t,P₂^t, . . . P_h^t]]+b_1t-1^c)
C_1t=f_1t-1*C_1t-1+i_1t-1*{tilde over (C)}_1t-1

σ( ) is a sigmoid function, tanh is an activation function, both W_1t-1ⁱand W_1t-1^care weight matrices, and both b_1t-1ⁱand b_1t-1^care offset values.

Finally, an output value h_1tis calculated based on the location features P₁^t, P₂^t, . . . , and P_h^tof the h target subjects in the image I_i, the output value h_1t-1and the input value C_1t:

o_1t-1=σ(W_1t-1^o·[h_1t-1,[P₁^t,P₂^t, . . . P_h^t]]+b_1t-1^o)
h_1t=o_1t-1tanh(C_1t)

σ( ) is a sigmoid function, tanh is an activation function, W_1t-1^ois a weight matrix, and b_1t-1^ois an offset value.

The foregoing h₁₁to h_1tmay constitute a first semantic description.

It may be understood that the initial output value h₁₀, offset values b₁₀^fto b_1t-1^f, offset values b₁₀ⁱto b_1t-1ⁱ, offset values b₁₀^cto b_1t-1^c, and offset values b₁₀^oto b_1t-1^omay be manually set. Weight matrices W₁₀^fto W_1t-1^f, weight matrices W₁₀ⁱto W_1t-1ⁱ, weight matrices W₁₀^cto W_1t-1^cmay be obtained by training a large quantity of known first semantic descriptions and location features of known target subjects.

The following provides calculation processes in a neuron 1, a neuron 2, . . . , and a neuron tin the second time sequence feature extraction unit.

In the neuron 1:

First, a forgotten value f₂₀is calculated based on attribute features S₁¹, S₂¹, . . . , and S_h¹of the h target subjects in the image I₁and an initial output value h₂₀:

f₂₀=σ(W₂₀^f·[h₂₀,[S₁¹,S₂¹, . . . S_h¹]]+b₂₀^f)

σ( ) is a sigmoid function, b₂₀^fis an offset value, and W₂₀^fis a weight matrix.

Then, an input value C₂₁is calculated based on the attribute features S₁¹, S₂¹, and S_h¹of the h target subjects in the image I₁, an initial input value C₂₀, the initial output value h₂₀, and the forgotten value f₂₀:

i₂₀=σ(W₂₀ⁱ·[h₂₀,[S₁¹,S₂¹, . . . S_h¹]]+b₂₀ⁱ)
{tilde over (C)}₂₀=tanh(W₂₀^c·[h₂₀,[S₁¹,S₂¹, . . . S_h¹]]+b₂₀^c)
C₂₁=f₂₀*C₂₀+i₂₀*{tilde over (C)}₂₀

σ( ) is a sigmoid function, tanh is an activation function, both W₂₀ⁱand W₂₀^care weight matrices, and both b₂₀ⁱand b₂₀^care offset values.

Finally, an output value h₂₁is calculated based on the attribute features S₁¹, S₂¹, . . . , and S_h¹of the h target subjects in the image I₁, the initial output value h₂₀and the input value C₂₁:

o₂₀=σ(W₂₀^o·[h₂₀,[S₁¹,S₂¹, . . . S_h¹]]+b₂₀^o)
h₂₁=o₂₀tanh(C₂₁)

σ( ) is a sigmoid function, tanh is an activation function, W₂₀^ois a weight matrix, and b₂₀^ois an offset value.

In the neuron 2:

First, a forgotten value f₂₁is calculated based on attribute features S₁², S₂², . . . , and S_h²of the h target subjects in the image I₂and the output value h₂₁:

f₂₁=σ(W₂₁^f·[h₂₁,[S₁²,S₂², . . . S_h²]]+b₂₁^f)

σ( ) is a sigmoid function, b₂₁^fis an offset value, and W₂₁^fis a weight matrix.

Then, an input value C₂₂is calculated based on the attribute features S₁², S₂², . . . , and S_h²of the h target subjects in the image I₂, an input value C₂₁, the output value h₂₁, and the forgotten value f₂₁:

i₂₁=σ(W₃₁ⁱ·[h₂₁,[S₁²,S₂², . . . S_h²]]+b₂₁ⁱ)
{tilde over (C)}₂₁=tanh(W₂₁^c·[h₂₁,[S₁²,S₂², . . . S_h²]]+b₂₁^c)
C₂₂=f₂₁*C₂₁+i₂₁*{tilde over (C)}₂₁

σ( ) is a sigmoid function, tanh is an activation function, both W₂₁ⁱand W₂₁^care weight matrices, and both b₂₁ⁱand b₂₁^care offset values.

Finally, an output value h₁₂is calculated based on the attribute features S₁², S₂², . . . , and S_h²of the h target subjects in the image I₂, the output value h₂₁and the input value C₂₂:

o₂₁=σ(W₂₁^o·[h₂₁,[S₁²,S₂², . . . S_h²]]+b₂₁^o)
h₁₂=o₁₁tanh(C₁₂)

σ( ) is a sigmoid function, tanh is an activation function, W₂₁^ois a weight matrix, and b₂₁^ois an offset value.

In the neuron t:

First, a forgotten value f_2t-1is calculated based on attribute features S₁^t, S₂^t, . . . , and S_h^tof the h target subjects in the image I_tand an output value h_2t-1:

f_2t-1=σ(W_2t-1^f·[h_2t-1,[S₁^t,S₂^t, . . . S_h^t]]+b_2t-1^f)

σ( ) is a sigmoid function, b_2t-1^fis an offset value, and W_2t-1^fis a weight matrix.

Then, an input value C_2tis calculated based on the attribute features S₁^t, S₂^t, . . . , and S_h^tof the h target subjects in the image I_t, an input value C_2t-1, the output value h_2t-1, and the forgotten value f_2t-1:

i_2t-1=σ(W_2t-1ⁱ·[h_2t-1,[S₁^t,S₂^t, . . . S_h^t]]+b_2t-1ⁱ)
{tilde over (C)}_2t-1=tanh(W_2t-1^c·[h_2t-1,[S₁^t,S₂^t, . . . S_h^t]]+b_2t-1^c)
C_2t=f_2t-1*C_2t-1+i_2t-1*{tilde over (C)}_2t-1

σ( ) is a sigmoid function, tanh is an activation function, both W_2t-1ⁱand W_2t-1^care weight matrices, and both b_2t-1ⁱand b_2t-1^cand are offset values.

Finally, an output value h_ais calculated based on the attribute features S₁^t, S₂^t, . . . , and S_h^tof the h target subjects in the image I_t, the output value h_2t-1and the input value C_2t:

o_2t-1=σ(W_2t-1^o·[h_2t-1,[S₁^t,S₂^t, . . . S_h^t]]+b_2t-1^o)
h_2t=o_2t-1tanh(C_2t)

σ( ) is a sigmoid function, tanh is an activation function, W_2t-1^ois a weight matrix, and b_2t-1^ois an offset value.

The foregoing h₂₁to h_2tmay constitute a second semantic description.

It may be understood that the initial output value h₂₀, offset values b₂₀^fto b_2t-1^f, offset values b₂₀ⁱto b_2t-1ⁱ, offset values b₂₀^cto b_2t-1^c, offset values b₂₀^oto b_2t-1^omay be manually set. Weight matrices W₂₀^fto W_2t-1^f, weight matrices W₂₀ⁱto W_2t-1ⁱ, weight matrices W₂₀^cto W_2t-1^care obtained by training a large quantity of known second semantic descriptions, known first semantic descriptions, and attribute features of known target subjects.

The following provides calculation processes in a neuron 1, neuron 2, . . . , and a neuron t in the third time sequence feature extraction unit.

In the neuron 1:

First, a forgotten value f₃₀is calculated based on posture features Z₁¹, Z₂¹, . . . , and Z_h¹of the h target subjects in the image I₁and an initial output value h₃₀:

f₃₀=σ(W₃₀^f·[h₃₀,[Z₁¹,Z₂¹, . . . Z_h¹]]+b₃₀^f)

σ( ) is a sigmoid function, b₃₀^fis an offset value, and W₃₀^fis a weight matrix.

Then, an input value C₃₁is calculated based on the posture features Z₁¹, Z₂¹, . . . , and Z_h¹of the h target subjects in the image I₁, an initial input value C₃₀, the initial output value h₃₀, and the forgotten value f₃₀:

i₃₀=σ(W₃₀ⁱ·[h₃₀,[Z₁¹,Z₂¹, . . . Z_h¹]]+b₃₀ⁱ)
{tilde over (C)}₃₀=tanh(W₃₀^c·[h₃₀,[Z₁¹,Z₂¹, . . . Z_h¹]]+b₃₀^c)
C₃₁=f₃₀*C₃₀+i₃₀*{tilde over (C)}₃₀

σ( ) is a sigmoid function, tanh is an activation function, both W₃₀ⁱand W₃₀^care weight matrices, and both b₃₀ⁱand b₃₀^care offset values.

Finally, an output value h₃₁is calculated based on the posture features Z₁¹, Z₂¹, . . . , and Z_h¹of the h target subjects in the image I₁, the initial output value h₃₀and the input value C₃₁:

o₃₀=σ(W₃₀^o·[h₃₀,[Z₁¹,Z₂¹, . . . Z_h¹]]+b₃₀^o)
h₃₁=o₃₀tanh(C₃₁)

σ( ) is a sigmoid function, tanh is an activation function, W₃₀^ois a weight matrix, and b₃₀^ois an offset value.

In the neuron 2:

First, a forgotten value f₃₁is calculated based on posture features Z₁², Z₂², . . . , and Z_h²of the h target subjects in the image I₂and the output value h₃₁:

f₃₁=σ(W₃₁^f·[h₃₁,[Z₁²,Z₂², . . . Z_h²]]+b₃₁^f)

σ( ) is a sigmoid function, b₃₁^fis an offset value, and W₃₁^fis weight matrix.

Then, an input value C₃₂is calculated based on the posture features Z₁², Z₂², . . . , and Z_h²of the h target subjects in the image I₂, an input value C₃₁, the output value h₃₁, and the forgotten value f₃₁:

i₃₁=σ(W₃₁ⁱ·[h₃₁,[Z₁²,Z₂², . . . Z_h²]]+b₃₁ⁱ)
{tilde over (C)}₃₁=tanh(W₃₁^c·[h₃₁,[Z₁²,Z₂², . . . Z_h²]]+b₃₁^c)
C₃₂=f₃₁*C₃₁+i₃₁*{tilde over (C)}₃₁

σ( ) is a sigmoid function, tanh is an activation function, both W₃₁ⁱand W₃₁^care weight matrices, and both b₃₁ⁱand b₃₁^care offset values.

Finally, an output value h₃₂is calculated based on the posture features Z₁², Z₂², . . . , and Z_h²of the h target subjects in the image I₂, the output value h₃₁and the input value C₃₂:

o₃₁=σ(W₃₁^o·[h₃₁,[Z₁²,Z₂², . . . Z_h²]]+b₃₁^o)
h₃₂=o₃₁tanh(C₃₂)

σ( ) is a sigmoid function, tanh is an activation function, W₃₁^ois a weight matrix, and b₃₁^ois an offset value.

In the neuron t:

First, a forgotten value f_3t-1is calculated based on posture features Z₁^t, Z₂^t, . . . , and Z_h^tof the h target subjects in the image I_tand an output value h_3t-1:

f_3t-1=σ(W_3t-1^f·[h_3t-1,[Z₁^t,Z₂^t, . . . Z_h^t]]+b_3t-1^f)

σ( ) is a sigmoid function, b_3t-1^fis an offset value, and W_3t-1^fis a weight matrix.

Then, an input value C_3tis calculated based on the posture features Z₁^t, Z₂^t, . . . , and Z_h^tof the h target subjects in the image I_t, an input value C_3t-1, the output value h_3t-1, and the forgotten value f_3t-1:

i_3t-1=σ(W_3t-1ⁱ·[h_3t-1,[Z₁^t,Z₂^t, . . . Z_h^t]]+b_3t-1ⁱ)
{tilde over (C)}_3t-1=tanh(W_3t-1^c·[h_3t-1,[Z₁^t,Z₂^t, . . . Z_h^t]]+b_3t-1^c)
C_3t=f_3t-1*C_3t-1+i_3t-1*{tilde over (C)}_3t-1

σ( ) is a sigmoid function, tanh is an activation function, both W_3t-1ⁱand W_3t-1^care weight matrices, and both b_3t-1ⁱand b_3t-1^care offset values.

Finally, an output value h_3tis calculated based on the posture features Z₁^t, Z₂^t, . . . , and Z_h^tof the h target subjects in the image I_t, the output value h_3t-1and the input value C_3t:

o_3t-1=σ(W_3t-1^o·[h_3t-1,[Z₁^t,Z₂^t, . . . Z_h^t]]+b_3t-1^o)
h_3t=o_3t-1tanh(C_3t)

σ( ) is a sigmoid function, tanh is an activation function, W_3t-1^ois a weight matrix, and b_3t-1^ois an offset value.

The foregoing h₃₁to h_3tmay constitute a third semantic description.

It may be understood that the initial output value h₃₀, offset values b₃₀^fto b_3t-1^f, offset values b₃₀ⁱto b_3t-1ⁱ, offset values b₃₀^cto b_3t-1, offset values b₃₀^oto b_3t-1^omay be manually set. Weight matrices W₃₀^fto W_3t-1^f, weight matrices W₃₀ⁱto W_3t-1ⁱ, weight matrices W₃₀^cto W_3t-1^care obtained by training a large quantity of known third semantic descriptions, known second semantic descriptions, and posture features of known target subjects.

The following provides calculation processes in a neuron 1, a neuron 2, . . . , and a neuron t in the fourth time sequence feature extraction unit.

In the neuron 1:

First, a forgotten value f₄₀is calculated based on relational vector features G₁₂¹, G₁₃¹, . . . , and G_h-1h¹of the h target subjects in the image I₁and an initial output value h₄₀:

f₄₀=σ(W₄₀^f·[h₄₀,[G₁₂¹,G₁₃¹, . . . G_h-1h¹]]+b₄₀^f)

σ( ) is a sigmoid function, b₄₀^fis an offset value, and W₄₀^fis a weight matrix.

Then, an input value C₄₁is calculated based on the relational vector features G₁₂¹, G₁₃¹, . . . , and G_h-1h¹of the h target subjects in the image I₁, an initial value C₄₀, the initial output value h₄₀, and the forgotten value f₄₀:

i₄₀=σ(W₄₀ⁱ,[h₄₀,[G₁₂¹,G₁₃¹, . . . G_h-1h¹]]+b₄₀ⁱ)
{tilde over (C)}₄₀=tanh(W₄₀^c·[h₄₀,[G₁₂¹,G₁₃¹, . . . G_h-1h¹]]+b₄₀^c)
C₄₁=f₄₀*C₄₀+i₄₀*{tilde over (C)}₄₀

σ( ) is a sigmoid function, tanh is an activation function, both W₄₀ⁱand W₄₀^care weight matrices, and both b₄₀ⁱand b₄₀^care offset values.

Finally, an output value h₄₁is calculated based on the relational vector features G₁₂¹, G₁₃¹, . . . , and G_h-1h¹of the h target subjects in the image I₁, the initial output value h₄₀and the input value C₄₁:

o₄₀=σ(W₄₀^o·[h₄₀,[G₁₂¹,G₁₃¹, . . . G_h-1h¹]]+b₄₀^o)
h₄₁=o₄₀tanh(C₄₁)

σ( ) is a sigmoid function, tanh is an activation function, W₄₀^ois a weight matrix, and b₄₀^ois an offset value.

In the neuron 2:

First, a forgotten value f₄₁is calculated based on relational vector features G₁₂², G₁₃², . . . , G_h-1h²of the h target subjects in the image I₂and the output value h₄₁:

f₄₁=σ(W₄₁^f·[h₄₁,[G₁₂²,G₁₃², . . . G_h-1h²]]+b₄₁^f)

σ( ) is a sigmoid function, b₄₁^fis an offset value, and W₄₁^fis a weight matrix.

Then, an input value C₄₂is calculated based on the relational vector features G₁₂², G₁₃², . . . , and G_h-1h²of the h target subjects in the image I₂, and input value C₄₁, the output value h₄₁, and the forgotten value f₄₁:

i₄₁=σ(W₄₁ⁱ·[h₄₁,[G₁₂²,G₁₃², . . . G_h-1h²]]+b₄₁ⁱ)
{tilde over (C)}₄₁=tanh(W₄₁^c·[h₄₁,[G₁₂²,G₁₃², . . . G_h-1h²]]+b₄₁^c)
C₄₂=f₄₁*C₄₁+i₄₁*{tilde over (C)}₄₁

σ( ) is a sigmoid function, tanh is an activation function, both W₄₁ⁱand W₄₁^care weight matrices, and both b₄₁ⁱand b₄₁^care offset values.

Finally, an output value h₄₂is calculated based on the relational vector features G₁₂², G₁₃², . . . , and G_h-1h²of the h target subjects in the image I₂, the output value h₄₁and the input value C₄₂:

o₄₁=σ(W₄₁^o·[h₄₁,[G₁₂²,G₁₃², . . . G_h-1h²]]+b₄₁^o)
h₄₂=o₄₁tanh(C₄₂)

σ( ) is a sigmoid function, tanh is an activation function, W₄₁^ois a weight matrix, and b₄₁^ois an offset value.

In the neuron t:

First, a forgotten value f_4t-1is calculated based on relational vector features G₁₂^t, G₁₃^t, . . . , and G_h-1h^tof the h target subjects in the image I_tand an output value h_4t-1:

f_4t-1=σ(W_4t-1^f·[h_4t-1,[G₁₂^t,G₁₃^t, . . . G_h-1h^t]]+b_4t-1^f)

σ( ) is a sigmoid function b_4t-1^fis an offset value, and W_4t-1^fis a weight matrix.

Then, an input value C_4tis calculated based on the relational vector features G₁₂^t, G₁₃^t, . . . , G_h-1h^tof the h target subjects in the image I_t, an input value C_4t-1, the output value h_4t-1, and the forgotten value f_4t-1:

i_4t-1=σ(W_4t-1ⁱ·[h_4t-1,[G₁₂^t,G₁₃^t, . . . G_h-1h^t]]+b_4t-1ⁱ)
{tilde over (C)}_4t-1=tanh(W_4t-1^c·[h_4t-1,[G₁₂^t,G₁₃^t, . . . G_h-1h^t]]+b_4t-1^c)
C_4t=f_4t-1*C_4t-1+i_4t-1*{tilde over (C)}_4t-1

σ( ) is a sigmoid function, tanh is an activation function, both W_4t-1ⁱand W_4t-1^care weight matrices, and both b_4t-1ⁱand b_4t-1^care offset values.

Finally, an output value h_4tis calculated based on the relational vector features G₁₂^t, G₁₃^t, . . . , and G_h-1h^tof the h target subjects in the image I_t, the output value h_4t-1and the input value C_4t:

o_4t-1=σ(W_4t-1^o·[h_4t-1,[G₁₂^t,G₁₃^t, . . . G_h-1h^t]]+b_4t-1^o)
h_4t=o_4t-1tanh(C_4t)

σ( ) is a sigmoid function, tanh is an activation function, W_4t-1^ois a weight matrix, and b_4t-1^ois an offset value.

The foregoing h₄₁to h_4tmay constitute a panoramic semantic description.

It may be understood that the initial output value h₄₀, offset values b₄₀^fto b_4t-1, offset values b₄₀ⁱto b_4t-1ⁱ, offset values b₄₀^cto b_4t-1^c, offset values b₄₀^oto b_4t-1^omay be manually set. Weight matrices W₄₀^fto W_4t-1^f, weight matrices W₄₀ⁱto W_4t-1ⁱ, weight matrices W₄₀^cto W_4t-1^care obtained by training a large quantity of known panoramic semantic descriptions, known third semantic descriptions, and relational vector features of known target subjects.

FIG. 7 is a schematic diagram of a flowchart of an image analysis method. The image analysis method in this implementation may include the following steps:

S101: An image analysis system obtains influencing factors oft frames of images, where the influencing factors include self-owned features of h target subjects in each of the t frames of images and relational vector features between the h target subjects in each of the t frames of images, and the self-owned features of each target subject include a location feature, an attribute feature, and a posture feature, where t and h are natural numbers greater than 1.

In an exemplary embodiment, the location feature is used to indicate a location of a corresponding target subject in the image. The location feature may be represented as (x, y, w, h), where x and y are a horizontal coordinate and a vertical coordinate of a central point of the target subject in the image, w is a width of the target subject in the image, and h is a height of the target subject in the image.

In an exemplary embodiment, the attribute feature may include a plurality of types of attribute features. Attribute features may be different if target subjects are different. For example, the target subject is a person. The attribute feature of the target subject may include one or more of a gender, a hairstyle, a clothes type, a clothes color, a height, a body size, and the like.

In an exemplary embodiment, the target subject also has a plurality of posture features. Posture features may be different if target subjects are different. For example, the target subject is a person. The posture feature of the target subject may include one or more of falling down, lying down, walking, running, jumping, and the like.

In an exemplary embodiment, the relational feature vector is a vector representing a relationship between two target subjects.

S102: The image analysis system obtains a panoramic semantic description based on the influencing factors.

In an exemplary embodiment, a panoramic semantic model reflects a mapping relationship between the influencing factors and the panoramic semantic description. The panoramic semantic model may be represented as:

y=Panorama(x).

x is the influencing factors of the panoramic semantic description, y is the panoramic semantic description, Panorama( ) is the mapping relationship between the influencing factors of the panoramic semantic description and the panoramic semantic description, and Panorama( ) may be obtained by training a large quantity of influencing factors of a known panoramic semantic description and the known panoramic semantic description.

In an exemplary embodiment, the panoramic semantic description can describe relationships between the target subject, relationships between the target subjects and actions, and relationships between the actions.

In an exemplary embodiment, feature extraction is performed on the t frames of images to obtain t feature vectors. Location feature extraction is performed on the t feature vectors to obtain the location features. Attribute feature extraction is performed on the t feature vectors to obtain the attribute features. Posture feature extraction is performed on the t feature vectors to obtain the posture features. Relational vector feature extraction is performed on the t feature vectors to obtain the relational vector features.

In an exemplary embodiment, extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features may be implemented separately by different CNN, or may be integrated into a same CNN. When the extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features is integrated into the same CNN, the extraction of the feature vectors, the location features, the attribute features, the posture features, and the relational vector features may be implemented at one layer in the CNN.

In an exemplary embodiment, a first semantic description is extracted based on the location features. A second semantic description is extracted based on the attribute features and the first semantic description. A third semantic description is extracted based on the posture features and the second semantics. The panoramic semantic description is extracted based on the relational vector features and the third semantic description.

In an exemplary embodiment, the first semantic description to the third semantic description, and the panoramic semantic description may be extracted by different RNN, or may be integrated into a same RNN for extraction. The RNN may include a LSTM model, BiLSTM model, and the like. When the first semantic description to the third semantic description, and the panoramic semantic description are integrated into the same RNN for extraction, the first semantic description to the third semantic description, and the panoramic semantic description may be extracted separately at different layers in the RNN.

For ease of description, definitions of the image, the target subject, the panoramic semantic description, and the like are not described in detail in this embodiment. For details, refer to FIG. 2 and FIG. 3 and descriptions of related definitions of the image, the target subject, the panoramic semantic model, the panoramic semantic description, and the like. The feature vectors, the location features, the attribute features, the posture features, the relational vector features, and extraction manners thereof are not described either in this embodiment. For details, refer to FIG. 4 and related descriptions. In addition, the panoramic semantic model and how to use the panoramic semantic model to perform the panoramic semantic description on the image are not described in detail in this embodiment. For details, refer to FIG. 5, FIG. 6A and FIG. 6B, and related descriptions.

In the solution, the higher-level panoramic semantic description may be obtained based on location features, attribute features, and posture features of a plurality of target subjects in a plurality of frames of images and relational vector features between the plurality of target subjects in the plurality of frames of images, to better reflect relationships between the plurality of subjects, between the subjects and actions, and between the actions in the images.

FIG. 8 is a schematic diagram of a structure of an image analysis system according to an implementation. The image analysis system in this embodiment includes a feature extraction module 510 and a panoramic semantic description module 520. The feature extraction module 510 includes a feature vector extraction unit 511, a location feature extraction unit 512, an attribute feature extraction unit 513, a posture feature extraction unit 514, and a relational vector feature unit 515. The panoramic semantic descriptor module 520 includes a first time sequence feature extraction unit 522, a second time sequence feature extraction unit 523, a third time sequence feature extraction unit 524, and a fourth time sequence feature extraction unit 525.

The feature extraction module 510 is configured to obtain influencing factors of a panoramic semantic description. The influencing factors include self-owned features of h target subjects in each of t frames of images and relational vector features between the h target subjects in each of the t frames of images, and the self-owned features include a location feature, an attribute feature, and a posture feature, where t and h are natural numbers greater than 1. The location feature is used to indicate a location of a corresponding target subject in the image. The attribute feature is used to indicate an attribute of the corresponding target subject. The posture feature is used to indicate an action of the corresponding target subject. The relational vector features are used to indicate relationships between target subjects.

The panoramic semantic description module 520 is configured to input the influencing factors into a panoramic semantic model to obtain the panoramic semantic description. The panoramic semantic model reflects a mapping relationship between the influencing factors and the panoramic semantic description. The panoramic semantic description can describe the relationships between target subjects, relationships between the target objects and actions, and the relationship between the actions.

In an exemplary embodiment, the location feature is used to indicate the location of the corresponding target subject in the image. The location feature may be represented as (x, y, w, h), where x and y are a horizontal coordinate and a vertical coordinate of a central point of the target subject in the image, w is a width of the target subject in the image, and h is a height of the target subject in the image.

In an exemplary embodiment, a relational feature vector is a vector representing a relationship between two target subjects.

In an exemplary embodiment, the panoramic semantic model reflects the mapping relationship between the influencing factors and the panoramic semantic description. The panoramic semantic model may be represented as:

y=Panorama(x).

In an exemplary embodiment, the feature vector extraction unit 511 is configured to extract features of the t frames of images to obtain t feature vectors. The location feature extraction unit 512 is configured to extract location features of the t feature vectors to obtain the location features. The attribute feature extraction unit 513 is configured to extract attribute features of the t feature vectors to obtain the attribute features. The posture feature extraction unit 514 is configured to extract posture features of the t feature vectors to obtain the posture features. The relational vector feature unit 515 is configured to extract relational vector features of the t feature vectors to obtain the relational vector features.

In an exemplary embodiment, the feature extraction module 510 includes a CNN. The feature vector extraction unit 511, the location feature extraction unit 512, the attribute feature extraction unit 513, the posture feature extraction unit 514, and the relational vector feature extraction unit 515 are integrated into the CNN. The feature vector extraction unit 511, the location feature extraction unit 512, the attribute feature extraction unit 513, the posture feature extraction unit 514, and the relational vector feature extraction unit 515 may be in different CNN, or may be integrated into a same CNN. The CNN may include a VGGNet, a ResNet, an FPNet, and the like. When the feature vector extraction unit 511, the location feature extraction unit 512, the attribute feature extraction unit 513, the posture feature extraction unit 514, and the relational vector feature extraction unit 515 are integrated into the same CNN, the feature vector extraction unit 511, the location feature extraction unit 512, the attribute feature extraction unit 513, the posture feature extraction unit 514, and the relational vector feature 515 may be a layer in the CNN.

In an exemplary embodiment, the first time sequence feature extraction unit 522 is configured to extract a first semantic description based on the location features. The second time sequence feature extraction unit is configured to extract a second semantic description based on the attribute features and the first semantic description. The third time sequence feature extraction unit is configured to extract a third semantic description based on the posture features and the second semantics. The fourth time sequence feature extraction unit is configured to extract the panoramic semantic description based on the relational vector features and the third semantic description.

In an exemplary embodiment, the panoramic semantic model 522 includes a RNN. The first time sequence feature extraction unit, the second time sequence feature extraction unit, the third time sequence feature extraction unit, and the fourth time sequence feature extraction unit are respectively one layer in the RNN. The first time sequence feature extraction unit to the fourth time sequence feature extraction unit may be in different RNN, or may be integrated into a same RNN. The RNN may include a LSTMmodel, a BiLSTM model, and the like. When the first time sequence feature extraction unit to the fourth time sequence feature extraction unit are integrated into the same RNN, the first time sequence feature extraction unit to the fourth time sequence feature extraction unit may be respectively a layer in the RNN.

In the solution, the higher-level panoramic semantic description can be obtained based on location features, attribute features, and posture features of a plurality of target subjects in a plurality of frames of images and relational vector features between the plurality of target subjects in the plurality of frames of images, to better reflect relationships between the plurality of subjects, between the subjects and actions, and between the actions in the images.

The image analysis system may be implemented on a compute node, or may be implemented on a cloud compute infrastructure. The following describes how to implement an image analysis system on a compute node and a cloud compute infrastructure.

As shown in FIG. 9, a compute node 100 may include a processor 110 and a memory 120. The processor is configured to run a feature extraction module 111 and a panoramic semantic model 112. The memory 120 is configured to store semantic descriptions, features, images, and the like 121. The compute node 100 further provides two external interface windows: a management interface 140 oriented to maintenance staff of a semantic description system and a user interface 150 oriented to a user. The interface window may have various forms, such as a web interface, a command line tool, and a REST interface.

In an exemplary embodiment, the management interface is used by the maintenance staff to input a large quantity of images used for a panoramic semantic description, a large quantity of known panoramic semantic descriptions, known third semantic descriptions, and relational vector features of known target subjects, a large quantity of known third semantic descriptions, known second semantic descriptions, and posture features of known target subjects, a large quantity of known second semantic descriptions, known first semantic descriptions, and attribute features of known target subjects, a large quantity of first semantic descriptions and location features of known target subjects to train the panoramic semantic model.

In an exemplary embodiment, the user interface is used by the user to input images whose panoramic semantic description needs to be extracted. The panoramic semantic description is output to the user through the user interface.

It may be understood that the compute node 100 is merely an example provided in this embodiment, and the compute node 100 may include more or fewer components than shown components, may combine two or more components, or may have different component configurations.

As shown in FIG. 10, a cloud compute infrastructure may be a cloud service cluster 200. The cloud service cluster 200 includes nodes and a communications network between the nodes. The node may be a compute node, or a virtual machine running on a compute node. The nodes may be classified into two types by function: a compute node 210 and a storage node 220. The compute node 210 is configured to run a feature extraction module 211 and a panoramic semantic model 212. The storage node 220 is configured to store semantic descriptions, features, images, and the like 221. The cloud service cluster 200 further provides two external interface windows: a management interface 240 oriented to maintenance staff of a question answering engine and a user interface 250 oriented to a user. The interface window may have various forms, such as a web interface, a command line tool, and a REST interface.

It may be understood that the cloud service cluster 200 is merely an example provided in this embodiment, and the cloud service cluster 200 may include more or fewer components than shown components, may combine two or more components, or may have different component configurations.

FIG. 11 is a schematic diagram of a structure of a semantic description system according to another embodiment. The semantic description system shown in FIG. 8 may be implemented in a compute node 300 shown in FIG. 11. The compute node 300 in this implementation includes one or more processors 311, a communications interface 312, and a memory 313. The processor 311, the communications interface 312, and the memory 313 may be connected through a bus 314.

The processor 311 includes one or more general purpose processors. The general purpose processor may be any type of device that can process an electronic instruction. The general purpose processor includes a central processing unit (CPU), a microprocessor, a microcontroller, a main processor, a controller, an application specific integrated circuit (ASIC), or the like. The processor 311 executes various types of digital storage instructions, for example, software or firmware instructions stored in the memory 313, so that the compute node 300 provides relatively wide range of services. For example, the processor 311 can execute a program or process data, to execute at least a part of the method discussed. The processor 311 may run the feature extraction module and the panoramic semantic model that are shown in FIG. 8.

The communications interface 312 may be a wired interface (for example, an Ethernet interface), and is configured to communicate with another compute node or user.

The memory 313 may include a volatile memory, for example, a random access memory (RAM). The memory may further include a non-volatile memory, for example, a read-only memory (ROM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD). Alternatively, the memory may include a combination of the foregoing types of memories. The memory 313 may store program code and program data. The program code includes feature extraction module code and panoramic semantic model code. The program data includes a large quantity of images used for a panoramic semantic description, a large quantity of known panoramic semantic descriptions, known third semantic descriptions, and relational vector features of known target subjects, a large quantity of known third semantic descriptions, known second semantic descriptions, and posture features of known target subjects, a large quantity of known second semantic descriptions, known first semantic descriptions, and attribute features of known target subjects, a large quantity of first semantic descriptions and location features of known target subjects, to train the panoramic semantic model.

The processor 311 may be configured to invoke the program code in the memory 313, to perform the following steps.

The processor 311 is configured to obtain influencing factors of t frames of images. The influencing factors include self-owned features of h target subjects in each of the t frames of images and relational vector features between the h target subjects in each of the t frames of images, and the self-owned features of each target subject include a location feature, an attribute feature, and a posture feature, where t and h are natural numbers greater than 1.

The processor 311 is configured to obtain a panoramic semantic description based on the influencing factors. The panoramic semantic description includes a description of relationships between target subjects, relationships between actions of the target subjects and the target subjects, and relationships between the actions of the target subjects.

FIG. 12 is a schematic diagram of a structure of a semantic description system according to still another embodiment. The semantic description system in this implementation may be implemented in a cloud service cluster shown in FIG. 12. The cloud service cluster includes at least one compute node 410 and at least one storage node 420.

The compute node 410 includes one or more processors 411, a communications interface 412, and a memory 413. The processor 411, the communications interface 412, and the memory 413 may be connected through a bus 417.

The processor 411 includes one or more general purpose processors. The general purpose processor may be any type of device that can process an electronic instruction. The general purpose processor includes a CPU, a microprocessor, a microcontroller, a main processor, a controller, an ASIC, or the like. The processor 411 can be a dedicated processor for the compute node 410 only or can be shared with other compute nodes 410. The processor 411 executes various types of digital storage instructions, for example, software or firmware instructions stored in the memory 413, so that the compute node 410 provides relatively wide range of services. For example, the processor 411 can execute a program or process data, to execute at least a part of the method discussed. The processor 411 may run the feature extraction module and the panoramic semantic model that are shown in FIG. 8.

The communications interface 412 may be a wired interface (for example, an Ethernet interface), and is configured to communicate with another compute node or user. When the communications interface 412 is the wired interface, the communications interface 412 may use a TCP/IP protocol suite, such as, an RAAS protocol, a remote function call (RFC) protocol, a simple object access protocol (SOAP) protocol, a simple network management protocol (SNMP) protocol, a common object request broker architecture (CORBA) protocol, and a distributed protocol.

The memory 413 may include a volatile memory, for example, a RAM. The memory may further include a non-volatile memory, for example, a ROM, a flash memory, a HDD), or a SSD. Alternatively, the memory may include a combination of the foregoing types of memories.

The compute node 420 includes one or more processors 421, a communications interface 422, and a memory 423. The processor 421, the communications interface 422, and the memory 423 may be connected through a bus 424.

The processor 421 includes one or more general purpose processors. The general purpose processor may be any type of device that can process an electronic instruction. The general purpose processor includes a CPU, a microprocessor, a microcontroller, a main processor, a controller, an ASIC, or the like. The processor 421 can be a dedicated processor for the storage node 420 only or can be shared with other storage nodes 420. The processor 421 executes various types of digital storage instructions, for example, software or firmware instructions stored in the memory 423, so that the storage node 420 provides relatively wide range of services. For example, the processor 421 can execute a program or process data, to execute at least a part of the method discussed.

The communications interface 422 may be a wired interface (for example, an Ethernet interface), and is configured to communicate with another compute device or user.

The storage node 420 includes one or more storage controllers 421 and a storage array 425. The storage controller 421 and the storage array 425 may be connected through a bus 426.

The storage controller 421 includes one or more general purpose processors. The general purpose processor may be any type of device that can process an electronic instruction. The general purpose processor includes a CPU, a microprocessor, a microcontroller, a main processor, a controller, an ASIC, or the like. The storage controller 421 can be a dedicated processor for the storage node 420 only or can be shared with the compute node 410 or other storage nodes 420. It may be understood that, in this embodiment, each storage node includes one storage controller. In another embodiment, a plurality of storage nodes may share one storage controller.

The storage array 425 may include a plurality of memories. The memory may be a non-volatile memory, such as a ROM, a flash memory, an HDD, or an SSD. The memory may also include a combination of the foregoing types of memories. For example, the storage array may include a plurality of HDDs or a plurality of SDDs, or the storage array may include an HDD and an SDD. With the assistance of the storage controller 421, the plurality of memories are combined in different manners to form a memory group, thereby providing higher storage performance and a data backup technology than a single memory. Optionally, the memory array 425 may include one or more data centers. The plurality of data centers may be disposed at a same location, or may be disposed separately at different locations. The storage array 425 may store program code and program data. The program code includes feature extraction module code and panoramic semantic model code. The program data includes a large quantity of images used for a panoramic semantic description, a large quantity of known panoramic semantic descriptions, known third semantic descriptions, and relational vector features of known target subjects, a large quantity of known third semantic descriptions, known second semantic descriptions, and posture features of known target subjects, a large quantity of known second semantic descriptions, known first semantic descriptions, and attribute features of known target subjects, a large quantity of first semantic descriptions and location features of known target subjects, to train a panoramic semantic model.

The compute node 410 may be configured to invoke the program code in the storage node 420, to perform the following steps.

The compute node 410 is configured to obtain influencing factors of t frames of images. The influencing factors include self-owned features of h target subjects in each of the t frames of images and relational vector features between the h target subjects in each of the t frames of images, and the self-owned features of each target subject include a location feature, an attribute feature, and a posture feature, where t and h are natural numbers greater than 1.

The compute node 410 is configured to obtain a panoramic semantic description based on the influencing factors. The panoramic semantic description includes a description of relationships between target subjects, relationships between actions of the target subjects and the target subjects, and relationships between the actions of the target subjects.

In the solution, the higher-level panoramic semantic description can be obtained based on location features, attribute features, and posture features of a plurality of target subjects in a plurality of frames of images and relational vector features between the plurality of target subjects in the plurality of frames of images, to better reflect relationships between the plurality of subjects, between the subjects and actions, and between the actions in the images.

All or some of the foregoing embodiments may be implemented by using software, hardware, firmware, or any combination thereof. When software is used to implement the embodiments, all or some of the embodiments may be implemented in a form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the procedure or functions according to the embodiments are all or partially generated. The computer may be a general purpose computer, a dedicated computer, a computer network, or other programmable apparatuses. The computer instructions may be stored in a computer-readable storage medium or may be transmitted from a computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website, computer, server, or data center to another website, computer, server, or data center in a wired (for example, a coaxial cable, an optical fiber, or a digital subscriber line) or wireless (for example, infrared, radio, or microwave) manner. The computer-readable storage medium may be any usable medium accessible by a computer, or a data storage device, such as a server or a data center, integrating one or more usable media. The usable medium may be a magnetic medium (for example, a floppy disk, a storage disk, or a magnetic tape), an optical medium (for example, a DVD), a semiconductor medium (for example, a SSD), or the like.

Number	Name	Date	Kind
20080193010	Eaton et al.	Aug 2008	A1
20170193286	Zhou	Jul 2017	A1
20180115706	Kang	Apr 2018	A1
20200019764	Sun	Jan 2020	A1
20210133461	Ren	May 2021	A1

Number	Date	Country
102880692	Jan 2013	CN
104244086	Dec 2014	CN
103345764	Feb 2016	CN
106650617	May 2017	CN
108197589	Jun 2018	CN
108416776	Aug 2018	CN
108509880	Sep 2018	CN
108510012	Sep 2018	CN
108960330	Dec 2018	CN
2993618	Mar 2016	EP

	Number	Date	Country
Parent	PCT/CN2019/107126	Sep 2019	WO
Child	17365089		US

Image analysis method and system

Information

Patent Number

Date Filed

Date Issued

Inventors

Original Assignees

Examiners

Agents

CPC

Field of Search

CPC

International Classifications

Term Extension

Abstract

Description

Claims

Priority Claims (1)

CROSS-REFERENCE TO RELATED APPLICATIONS

US Referenced Citations (5)

Foreign Referenced Citations (10)

Non-Patent Literature Citations (1)

Related Publications (1)

Continuations (1)