Back-propagation image visual saliency detection method based on depth image mining

Abstract

A back-propagation significance detection method based on depth map mining, comprising: for an input image Io, at a preprocessing phase, obtaining a depth image Id and an image Cb with four background corners removed of the image Io; at a first processing phase, carrying out positioning detection on a significant region of the image by means of the obtained image Cb with four background corners removed and the obtained depth image Id to obtain the preliminary detection result S1 of a significant object in the image; then carrying out depth mining on a plurality of processing phases of the depth image Id to obtain corresponding significance detection results; and then optimizing the significance detection result mined in each processing phase by means of a back-propagation mechanism to obtain a final significance detection result map. The method can improve the detection accuracy of the significance object.

Description

CROSS-REFERENCE TO RELATED PATENT APPLICATIONS

The present application is a U.S. national stage application under 35 U.S.C. § 371 of PCT Application No. PCT/CN2017/112788, filed Nov. 24, 2017, which claims priority to Chinese Patent Application No. 201710513077.2, filed Jun. 29, 2017. The disclosures of the aforementioned priority applications are incorporated herein by reference in their entireties.

TECHNICAL FIELD

The present invention relates to the technical field of image processing, in particular to a multi-phase back-propagation image visual saliency detection algorithm for deep mining of a depth image.

BACKGROUND ART

In a complex scene, human would be quickly attracted to a few of salient visual objects and process these objects preferentially with eyes. This process is called visual saliency. The saliency detection is just to simulate human eyes to properly process images with a mathematical calculation method by using a visually biological mechanism of the human eyes, so as to obtain a salient object in one picture. Since we can preferentially distribute computing resources required by image analysis and synthesis through salient regions, it is significant to detect the salient regions of the images by calculation. The extracted salient images can be widely used in many computer vision fields, including image segmentation of target objects of interest, detection and recognition of the target objects, image compression and encoding, image retrieval, content-aware image editing and the like.

Generally, an existing saliency detection framework is mainly divided into: a from-bottom-to-top saliency detection method and a from-top-to-bottom saliency detection method. The data driving-based from-bottom-to-top saliency detection method is adopted mostly at the present and independent from specific tasks. The from-top-to-bottom saliency detection method is subjected to consciousness and associated with specific tasks.

In the existing methods, most of the from-bottom-to-top saliency detection methods use low-level characteristic information, such as color characteristics, distance characteristics and some heuristic saliency characteristics. Although these methods have their own advantages, they are not accurate and robust enough on challenging data sets in specific scenes. In order to solve this problem, with the advent of a 3D (3-dimensional) image acquisition technology, there are methods adopting depth information to enhance the accuracy of salient object detection at the present. The depth information can increase the accuracy of the salient object detection, but when one salient object has a low contrast with its background, the accuracy of the saliency detection will still be affected. On the whole, the existing image salient object detection methods have low accuracy during detection of the salient objects, are not robust enough and may easily cause error detection, missing detection and the like, and it is very hard to obtain an accurate image saliency detection result, resulting in false detection of a salient object body and also causing a certain error to an application using a saliency detection result.

SUMMARY OF THE INVENTION

The present invention is directed to provide a back-propagation saliency detection algorithm for deep mining of a depth image for the abovementioned shortcomings in the prior art, so as to solve the problems that existing saliency detection is not accurate and robust enough, allow a salient region in an image to be displayed more accurately and provide accurate and useful information for later applications such as target recognition and classification.

A technical solution provided by the present invention is:

A back-propagation saliency detection method based on depth image mining, including: at a preprocessing phase, obtaining a depth image of an image and an image with four background corners removed; at a first processing phase, carrying out positioning detection on a salient region of the image by means of color, depth and distance information to obtain a preliminary detection result of a salient object in the image; then carrying out deep mining on the depth image from a plurality of layers (processing phases) to obtain corresponding saliency detection results; and then optimizing the saliency detection result mined in each layer by means of a back-propagation mechanism to obtain a final saliency detection result image. The implementation of the method includes the following steps:

1) A preprocessing phase: for an input image I_o, firstly obtaining a depth image, defined as I_d, by means of Kinect equipment; and secondly, removing four background edges of the image by means of a BSCA algorithm, and defining the obtained image with the four background corners removed as C_b, wherein the BSCA algorithm is recorded in the document (Qin Y, Lu H, Xu Y, et al. Saliency detection via Celluar Automata [C]/IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2015:110-119, and is to obtain a background saliency map based on background seed information according to color and distance information comparison.

2) A first processing phase: carrying out preliminary saliency detection on the input image I_o by means of the obtained image C_b with the four background corners removed and the depth image I_d to obtain a preliminary saliency monitoring result defined as: S₁;

specifically including steps 11 to 15:

Step 11, dividing the image into K regions by means of a K-means algorithm, and calculating a color saliency value S_c(r_k) of each subregion through the formula (1):S_c(r_k)=Σ_i=1,i≠k^KP_iW_s(r_k)D_c(r_k,r_i)  (1)wherein r_k and r_i respectively represent regions k and i, D_c(r_k, r_i) represents a Euclidean distance of the region k and the region i in an L*a*b color space, P_i represents a proportion of the region i to an image region; W_s(r_k) is defined as follows:

$\begin{array}{cc}{W}_s\left(r_k\right)=e^{-\frac{D_o\left(r_k,r_i\right)}{{\sigma }^2}}& \left(2\right)\end{array}$ wherein D_o(r_k, r_i) represents a coordinate position distance of the region k and the region i, and σ is a parameter controlling the range of the W_s(r_k).

Step 12, by means of the color saliency value calculating mode, calculating a depth saliency value S_d(r_k) of the depth image through the formula (3):

$\begin{array}{cc}{S}_d\left(r_k\right)=\sum _{i=1,i\ne k}^K{P}_i{W}_s\left(r_k\right)D_d\left(r_k,r_i\right)& \left(3\right)\end{array}$ wherein D_d(r_k, r_i) represents a Euclidean distance of the region k and the region i in a depth space. Step 13, calculating a center and a depth weight S_s(r_k) of the region k through the formula (4), wherein generally, a salient object is located in the center:

$\begin{array}{cc}{S}_s\left(r_k\right)=\frac{G\left(P_k-P_o\right)}{N_k}{W}_d\left(d_k\right)& \left(4\right)\end{array}$ wherein G(⋅) represents Gaussian normalization, ∥⋅∥ represents Euclidean distance operation, P_k is a position coordinate of the region k, P_o is a coordinate center of the image, and N_k is the number of pixels of the region k. W_d(d_k) is a depth weight, defined as follows:W_d(d_k)=(max{d}−d_k)^μ  (5)wherein max{d} represents a maximum depth of the depth image, d_k represents a depth value of the region k, and u is a parameter related to the calculated depth image, defined as follows:

$\begin{array}{cc}\mu =\frac{1}{\max \left\{d\right\}-\min \left\{d\right\}}& \left(6\right)\end{array}$ wherein min{d} represents a minimum depth of the depth image.

Step 14, obtaining a coarse saliency detection result S_fc(r_k) by means of the formula (7), which is a non-optimized preliminary saliency detection result at the first processing phase:S_fc(r_k)=G(S_c(r_k)S_s(r_k)+S_d(r_k)S_s(r_k))  (7)

Step 15, in order to optimize the preliminary saliency detection result, enhancing the result of the formula (7) by means of the depth image I_d(d_k) and the image C_b with the four background corners removed at the preprocessing phase, as shown in the formula 8:S₁(r_k)=s_fc(r_k)׬I_d(d_k)×c_b  (8)S₁(r_k) represents an optimized result of the S_fc(r_k) of the formula 7, namely an optimized detection result at the first processing phase;

3) A second processing phase: converting the depth image into a color image, and obtaining a medium saliency detection result, defined as S₂, by means of the calculating process of the first processing phase and optimization of the back-propagation mechanism.

3) A third processing phase: carrying out background filtering on the depth image, converting the filtered depth image into a color image, and obtaining a final saliency detection result S by means of the calculating process of the second processing phase and optimization of the back-propagation mechanism.

Compared with the prior art, the present invention has the beneficial effects that:

the present invention provides the multi-layer back-propagation saliency detection algorithm based on depth image mining, including: firstly, at a preprocessing phase, obtaining the depth image of the image and the image with four background corners removed; secondly, calculating the preliminary saliency detection result based on information such as the color, the space and the depth of the image by means of the saliency detection algorism of the first layer; then carrying out deep mining on the depth image from the second layer and the third layer, and carrying out saliency detection by means of the calculating mode of the first layer; and finally, optimizing the saliency detection results of the second and third layers by means of the back-propagation mechanism to obtain the secondary saliency detection result image and the final saliency detection result image.

The present invention can detect the salient object more accurately and robustly. Compared with the prior art, the present invention has the following technical advantages:

(I) by multi-layer mining of the depth image, the present invention can improve the accuracy of salient object detection; and

(II) the present invention provides the back-propagation mechanism to optimize the saliency detection result of each layer.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a flowchart provided by the present invention.

FIG. 2 shows a comparison diagram of detection result images, obtained by detecting an input image by respectively adopting an existing method and the method of the present invention, and images expected to be obtained via artificial calibration according to an embodiment of the present invention, wherein the first column displays the input images; the second column displays the images expected to be obtained via the artificial calibration; the third column displays the detection result images of the present invention; and the columns from four to ten are detection result images obtained by means of other existing methods.

DETAILED DESCRIPTION OF THE INVENTION

The present invention is further described below through embodiments in combination with drawings, but the scope of the present invention will not be limited in any mode.

The present invention provides a multi-layer back-propagation saliency detection algorithm based on depth image mining, which can detect a salient object more accurately and robustly. In the present invention, firstly, at a preprocessing layer/phase, a depth image of an image and an image with four background corners removed are obtained; secondly, at a first layer, second layer and third layer, the depth image is mined respectively to obtain corresponding saliency detection results; and finally, a back-propagation mechanism is used to optimize the saliency detection results of the various layers/processing phases to obtain a final saliency detection result image. FIG. 1 is a flowchart of a salient object detection method provided by the present invention, including the following steps that:

Step I, an image I_o to be detected is input, and an image C_b with four background corners removed and a depth image I_d of the image are obtained;

wherein the image C_b with four background corners removed is obtained by means of a BSCA algorithm recorded in the document (Qin Y, Lu H, Xu Y, et al. Saliency detection via Celluar Automata [C]/IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2015:110-119, and the depth image I_d of the image is obtained by means of Kinect equipment;first layer of operation of the algorithm: (Steps II to VI)

Step II, the image is divided into K regions by means of a K-means algorithm, and a color saliency value S_c(r_k) of each subregion is calculated through the formula (1):S_c(r_k)=Σ_i=1,i≠k^KP_iW_s(r_k)D_c(r_k,r_i)  (1)wherein r_k and r_i respectively represent regions k and i, D_c(r_k, r_i) represents a Euclidean distance of the region k and the region i in an L*a*b color space, P_i represents a proportion of the region i to an image region; W_s(r_k) is defined as follows:

$\begin{array}{cc}{W}_s\left(r_k\right)=e^{-\frac{D_o\left(r_k,r_i\right)}{{\sigma }^2}}& \left(2\right)\end{array}$ wherein D_o(r_k, r_i) represents a coordinate position distance of the region k and the region i, and σ is a parameter controlling the range of the W_s(r_k);

Step III, by means of the color saliency value calculating mode, a depth saliency value S_d(r_k) of the depth image is calculated through the formula (3):

$\begin{array}{cc}{S}_d\left(r_k\right)=\sum _{i=1,i\ne k}^K{P}_i{W}_s\left(r_k\right)D_d\left(r_k,r_i\right)& \left(3\right)\end{array}$ wherein D_d(r_k, r_i) represents a Euclidean distance of the region k and the region i in a depth space. Step IV, calculating a center and a depth weight S_s(r_k) through the formula (4), wherein generally, a salient object is located in the center:

$\begin{array}{cc}{S}_s\left(r_k\right)=\frac{G\left(P_k-P_o\right)}{N_k}{W}_d\left(d_k\right)& \left(4\right)\end{array}$ wherein G(⋅) represents Gaussian normalization, ∥⋅∥ represents Euclidean distance operation, P_k is a position coordinate of the region k, P_o is a coordinate center of the image, and N_k is the number of pixels of the region k. W_d(d_k) is a depth weight, defined as follows:W_d(d_k)=(max{d}−d_k)^μ  (5)wherein max{d} represents a maximum depth of the depth image, d_k represents a depth value of the region k, and u is a parameter related to the calculated depth image, defined as follows:

$\begin{array}{cc}\mu =\frac{1}{\max \left\{d\right\}-\min \left\{d\right\}}& \left(6\right)\end{array}$ wherein min{d} represents a minimum depth of the depth image;

Step V, a coarse saliency detection result S_fc(r_k) is obtained by means of the formula (7), which is a non-optimized preliminary saliency detection result at the first processing phase:S_fc(r_k)=G(S_c(r_k)S_s(r_k)+S_d(r_k)S_s(r_k))  (7);

Step VI, in order to optimize the preliminary saliency detection result, the result of the formula (7) is enhanced by means of the depth image I_d(d_k) and the image C_b with the four background corners removed at the preprocessing phase, as shown in the formula 8:S₁(r_k)=s_fc(r_k)׬I_d(d_k)×c_b  (8)

S₁(r_k) represents an optimized result of the S_fc(r_k) of the formula 7, namely an optimized detection result at the first processing phase;

second layer of operation of the algorithm: (Steps VII to IX)

Step VII, the depth image is further mined: firstly, the depth image is extended into a depth-based color image through the formula (9):I_eR|G|B=I_oR|G|B×I_d  (9)wherein I_e is the extended depth-based color image;

Step VIII, the operations in the steps II to V of the first layer are carried out on the extended depth-based color image to obtain a coarse saliency detection result S_sc(r_k):S_sc(r_k)=G(S_c(r_k)S_s(r_k)+S_d(r_k)S_s(r_k)),  (10);

Step IX, in order to optimize the coarse saliency detection result, a back-propagation mechanism is used to optimize the coarse saliency detection result by means of the preliminary detection result (the result calculated through the formula (7)) of the first layer through the formula (11), so as to obtain a medium saliency detection result S₂(r_k):S₂(r_k)=S₁²(r_k)(1−e^{−Ssc2(rk)-Id(rk)})  (11);third layer of operation of the algorithm: (Steps X to XIII)

Step X, the depth image is further mined: firstly, background filtering processing is carried out on the depth image by means of the formula (12) to obtain a filtered depth image I_df:

$\begin{array}{cc}{I}_{\mathrm{df}}=\{\begin{array}{cc}{I}_d,& d\le \beta \times \max \left\{d\right\}\\ 0,& d>\beta \times \max \left\{d\right\}\end{array}& \left(12\right)\end{array}$ wherein I_df is the depth image with the background filtered;

Step XI, the filtered depth image is extended into a color image, defined as I_ef, through the operation of the formula (9) of the step VII of the second layer;

Step XII, the color image I_ef of the filtered depth image is operated through the steps II to V of the first layer to obtain a coarse saliency detection result S_tc(r_k) of the third layer:S_tc(r_k)=G(S_c(r_k)S_s(r_k)+S_d(r_k)S_s(r_k))  (13);

Step VIII, in order to optimize the coarse saliency detection result, the back-propagation mechanism is used to optimize the coarse detection result of the third layer by means of the preliminary detection results of the first layer and the second layer through the formula (14) to obtain a final saliency detection result S(r_k):S(r_k)=S₂(r_k)(S₂(r_k)+S_tc(r_k))(S_tc(r_k)+1−e^{−Stc2(rk)Sk(rk)})  (14)

FIG. 2 shows detection result images, obtained by detecting an input image by respectively adopting an existing method and the method of the present invention, and images expected to be obtained via artificial calibration, wherein the first column displays the input images; the second column displays the images expected to be obtained via the artificial calibration; the third column displays the detection result images of the present invention; and the columns from four to ten are detection result images obtained by means of other existing methods. Through comparison with the images of FIG. 2, it can be seen that compared with other methods, the method of the present invention can detect the salient object, is lowest in error rate and highest in accuracy and has extremely good robustness.

It should be noted that the embodiments are disclosed to help to further understand the present invention, but those skilled in the art can understand that various replacements and modifications are possible without departing from the spirit and scope of the present invention and attached clamps. Therefore, the present invention should not be limited to the contents disclosed by the embodiments. The protection scope of the present invention is based on the scope defined by claims.

Claims

1. A back-propagation saliency detection method based on depth image mining, comprising, for an input image Io: at a preprocessing phase, obtaining a depth image Id of an image Io and an image Cb with four background corners removed;at a first processing phase, carrying out positioning detection on a salient region of the image by means of the obtained image Cb with four background corners removed and the depth image Id to obtain a preliminary detection result of a salient object in the image, comprising steps 11-14;Step 11, dividing the image into K regions by means of a K-means algorithm, and calculating a color saliency value Sc(rk) of each subregion through the formula (1):

2. The back-propagation saliency detection method based on depth image mining of claim 1, wherein the depth image Id is specifically obtained by means of Kinect equipment.