Patent Application
20230298445
The present invention relates to a learning apparatus, an estimation apparatus, a learning method, and a program.
Patent Document 1 discloses a technique for generating an estimation model for classifying an input image into a good image or a bad image by learning based on training images of a correct answer and an incorrect answer. A good image is an image having a high similarity with respect to a training image of a correct answer, and a bad image is an image having a low similarity with respect to a training image of a correct answer. Patent Document 2 discloses a technique for defining an abnormal behavior by a training image indicating an abnormal behavior, and generating an estimation model for detecting the defined abnormal behavior.
In a technique for generating an estimation model for detecting abnormality, a technique for efficiently collecting a training image has been desired. Patent Document 1 does not disclose the problem and a solving means. In a case of a technique described in Patent Document 2, it is necessary to collect a large number of training images indicating an abnormal behavior. However, it is not easy to collect a training image indicating “abnormality”. An object of the present invention is to provide a technique for efficiently collecting a training image for generating an estimation model for detecting abnormality.
The present invention provides a learning apparatus including:
Further, the present invention provides a learning method including, by a computer:
Further, the present invention provides a program causing a computer to function as:
Further, the present invention provides an estimation apparatus for discriminating between normal and abnormal by using an estimation model generated by the learning apparatus.
The present invention enables to efficiently collect a training image for generating an estimation model for detecting abnormality.
Hereinafter, example embodiments according to the present invention are described with reference to the drawings. Note that, in all drawings, a similar constituent element is indicated by a similar reference sign, and description thereof is omitted as necessary.
A learning apparatus according to a present example embodiment (hereinafter, may simply be referred to as a “learning apparatus”) generates an estimation model for discriminating whether a state indicated by an input image is normal or abnormal.
A discrimination target regarding normal and abnormal is, for example, a place (such as a park, a station, and an institution). A regular state being observed during most of time is discriminated to be normal, and a state different from the regular state is discriminated to be abnormal. For example, a state in which a person performing an abnormal behavior is present, a state in which an object always present at the place is out of order or has been moved, or the like is discriminated to be abnormal. The abnormal behavior is a behavior different from a behavior being performed by a majority of people being observed in an image. Note that, in addition to the above, the discrimination target may be a facility such as a factory, a store, an institution, and an office, or may be other than the above. In any case, a regular state being observed during most of time is discriminated to be normal, and a state different from the regular state is discriminated to be abnormal.
The learning apparatus generates the above-described estimation model by repeatedly performing a cycle illustrated in
The first image registration processing S1 is processing of classifying and registering an image generated by a camera, based on a similarity between the image generated by the camera, and an image being registered in advance and indicating an abnormal state.
First to third image group DBs 17-1 to 17-3, a camera D14, similarity computation S10, and registration S11 in
First, as pre-preparation of the processing, a labeled image attached with a label of an abnormal state is stored in the first image group database (DB) 17-1. A user prepares in advance several images indicating an abnormal state, and stores, in the first image group DB 17-1, the images by attaching a label of an abnormal state. Images accumulated in the first image group DB 17-1 as described above are labeled images being confirmed to indicate an abnormal state by a user, and having high reliability. Note that, images to be stored for the first time in the first image group DB 17-1 may be from several tens to several hundreds of images, and a large number of images are not necessary. The number of such degree as described above does not increase a user load required for collecting labeled images. Note that, in a case where an abnormal state is defined in advance, and an estimation model for detecting the abnormal state is generated, generally, it is necessary to prepare several thousands to several ten thousands of training images indicating an abnormal state. The first image group DB 17-1 is equivalent to the image storage unit 17 in
The acquisition unit 11 acquires an image generated by the camera D14. The camera D14 may be a camera (such as a surveillance camera) for photographing a discrimination target regarding normal and abnormal, or may be a camera for photographing a target of a same type as a discrimination target. The camera D14 may photograph a moving image, or may photograph a still image successively at a frame interval longer than that of a moving image. In
The acquisition unit 11 may acquire an image generated by the camera D14 by real-time processing. In this case, the learning apparatus 10 and the camera D14 are configured to be communicable with each other. In addition to the above, the acquisition unit 11 may acquire an image generated by the camera D14 by batch processing. In this case, an image generated by the camera D14 is accumulated in a storage apparatus included in the camera D14 or any other storage apparatus, and the acquisition unit 11 acquires the accumulated image at any timing.
Note that, in the present description, “acquisition” includes at least one of “acquisition of data stored in another apparatus or a storage medium by an own apparatus (active acquisition)”, based on a user input, or based on a command of a program, for example, requesting or inquiring another apparatus and receiving, accessing to another apparatus or a storage medium and reading, and the like, “input of data to be output from another apparatus to an own apparatus (passive acquisition)”, based on a user input, or based on a command of a program, for example, receiving data to be distributed (or transmitted, push-notified, or the like), and acquiring by selecting from received data or information, and “generating new data by editing data (such as converting into a text, rearranging data, extracting a part of pieces of data, and changing a file format) and the like, and acquiring the new data”.
The similarity computation unit 12 computes a similarity between an image (hereinafter, referred to as an “acquired image”) acquired by the acquisition unit 11, and a first image being accumulated in advance in the first image group DB 17-1 and indicating an abnormal state (S10 in
Note that, various methods have been proposed in computation of a similarity between images. In the present example embodiment, any method can be adopted. For example, the similarity computation unit 12 may detect an object from an image, and compute a similarity of a detection result (such as a similarity of the number of detected objects, and a similarity of an external appearance of a detected object). Further, the similarity computation unit 12 may input each image to an estimation model for analyzing an image generated by deep learning, and compute a similarity of an analysis result of an acquired image (such as a recognition result of an object indicated by an image, and a recognition result of a scene indicated by an image). Furthermore, the similarity computation unit 12 may compute a similarity of a color or a luminance appearing in an entirety or a local portion of an image.
The registration unit 13 registers, in the second image group database (DB) 17-2, an acquired image whose similarity is equal to or less than a first reference value, as a second image indicating a normal state (an image attached with a label of a normal state) (S11). In a case where the similarity computation unit 12 computes a similarity between each of a plurality of first images accumulated in the first image group DB 17-1, and each acquired image, the registration unit 13 registers, in the second image group DB 17-2, an acquired image whose similarity with respect to all of the plurality of first images is equal to or less than the first reference value, as a second image.
Further, the registration unit 13 registers, in the third image group database (DB) 17-3, an acquired image whose similarity is equal to or more than a second reference value, as a third image indicating an abnormal state (an image attached with a label of an abnormal state) (S11). In a case where the similarity computation unit 12 computes a similarity between each of a plurality of first images accumulated in the first image group DB 17-1, and each acquired image, the registration unit 13 registers, in the third image group DB 17-3, an acquired image whose similarity with respect to at least one of the plurality of first images is equal to or more than the second reference value, as a third image.
An image determined to be similar to a first image by a predetermine level or more by a computer as described above is registered in the third image group DB 17-3, as an image indicating an abnormal state. In this regard, the third image group DB 17-3 is different from the first image group DB 17-1 for storing a first image being confirmed to indicate an abnormal state by a user and having high reliability.
The first reference value and the second reference value may be a same value, or may be a different value. However, setting the first reference value and the second reference value to a different value from each other, setting the first reference value to a sufficiently small value, and setting the second reference value to a sufficiently large value enables to suppress an inconvenience that an acquired image being present in a gray zone (where a similarity is larger than the first reference value, and smaller than the second reference value) where a similarity to a first image is neither high nor low is registered as a second image or a third image.
The image selection processing S2 is processing of selecting an image to be set as a training image from among images accumulated in the first to third image group DBs 17-1 to 17-3. The learning processing S3 is processing of performing learning of each of a plurality of estimation models registered in an estimation model database (DB) 18-1, while using a selected image as a training image.
The first to third image group DBs 17-1 to 17-3, the estimation model DB 18-1, selection S12, and learning S13 in
First, information on a plurality of estimation models is stored in the estimation model DB 18-1. All of the plurality of estimation models are models for discriminating whether a state indicated by an input image is normal or abnormal. In the plurality of estimation models, a learning algorithm and an estimation algorithm are different from each other. For example, a plurality of estimation models are generated by deep learning. In the present example embodiment, for example, information on a plurality of estimation models learned and generated by a neural network, a Bayesian network, a regression analysis, a support vector machine (SVM), a decision tree, a genetic algorithm, a nearest neighbor classification method, and the like is stored in the estimation model DB 18-1.
The learning unit 14 selects at least a part of images from among images registered in the first to third image group DBs 17-1 to 17-3 (S12 in
Various selection methods are available. For example, the learning unit 14 may at random select a predetermined number of images determined in advance from the entirety of the first to third image group DBs 17-1 to 17-3. In addition to the above, the learning unit 14 may at random select a first predetermined number of images determined in advance from the first image group DB 17-1, may at random select a second predetermined number of images determined in advance from the second image group DB 17-2, and may at random select a third predetermined number of images determined in advance from the third image group DB 17-3. The first to third predetermined numbers may be a same number, or may be a different number. Specifically, a ratio (a ratio with respect to the entirety of images to be selected) of the number of images to be selected from each of the first to third image group DBs 17-1 to 17-3 may be the same, or may be different.
Further, the learning unit 14 may select an image for each estimation model. In this case, the above-described first to third predetermined numbers and the above-described ratio may be different for each estimation model.
After selecting an image, the learning unit 14 performs learning of each of a plurality of estimation models registered in the estimation model database (DB) 18-1, while using selected first to third images as training images. Specifically, the learning unit 14 generates an estimation model for discriminating between normal and abnormal by machine learning (a concept including deep learning) using first to third images.
The estimation processing S4 is processing of inputting an acquired image to each of a plurality of estimation models registered in the estimation model database (DB) 18-1, and discriminating a state indicated by the acquired image.
The estimation model DB 18-1, the camera D14, and estimation S14 in
The learning-time estimation unit 15 inputs an acquired image to each of a plurality of estimation models stored in the estimation model storage unit 18, and discriminates a state (normal or abnormal) indicated by the acquired image. Note that, an acquired image to be input to an estimation model by the processing is an acquired image being not used for generation (learning) of the estimation model at the point of time. For example, the learning-time estimation unit 15 can perform the discrimination by using an acquired image before being stored in the image storage unit 17.
Note that, a discrimination result of each of a plurality of estimation models may be accumulated in a storage apparatus in the learning apparatus 10.
The user confirmation processing S5 is processing of outputting, toward a user, a discrimination result in the estimation processing S4, and accepting, from the user, a correct/incorrect input of the discrimination result.
A display apparatus D15, extraction S15, output S16, and correct/incorrect input S17 in
The user confirmation unit 16 outputs, toward a user, a discrimination result by the learning-time estimation unit 15 (S16 in
When the processing is performed with respect to all the acquired images, a load of a user may increase. In view of the above, the user confirmation unit 16 may extract a part of acquired images that satisfy a predetermined condition (S15 in
A part of acquired images for which an output of a discrimination result and an acceptance of a correct/incorrect input are performed may be, for example, any one of the following images.
A part of acquired images for which an output of a discrimination result and an acceptance of a correct/incorrect input are performed may include, in addition to any one of the above-described acquired images, an acquired image picked up at random from among acquired images (acquired images presumed to indicate a normal state) that do not satisfy the above-described condition.
The user confirmation unit 16 may perform an output of a discrimination result via any output apparatus such as a display or a projection apparatus, and accept a correct/incorrect input via any input apparatus such as a keyboard, a mouse, a touch panel, a physical button, or a microphone. In addition to the above, the user confirmation unit 16 may transmit a discrimination result to a predetermined mobile terminal, and acquire, from the mobile terminal, a content of a correct/incorrect input performed for the mobile terminal. In addition to the above, the user confirmation unit 16 may store the discrimination result in any server in a state browsable from any apparatus. Further, the user confirmation unit 16 may acquire a content of a correct/incorrect input being input from any apparatus and stored in the above-described server. Note that, the example described herein is merely one example, and the present example embodiment is not limited thereto.
The second image registration processing S6 is processing of registering, in the first image group DB 17-1, an acquired image being input indication of an abnormal state in the user confirmation processing S5, as a first image.
The first image group DB 17-1 and registration S18 in
The registration unit 13 registers, in the first image group DB 17-1, an acquired image being input indication of an abnormal state in a correct/incorrect input to be accepted by the user confirmation unit 16, as a first image.
The acquired image being input indication of an abnormal state corresponds to an acquired image in which a discrimination result is “abnormal state” and a correct/incorrect input is “correct”, an acquired image in which a discrimination result is “normal state” and a correct/incorrect input is “incorrect”, and the like.
Herein, a modification example of the learning apparatus 10 according to the present example embodiment is described. The learning apparatus 10 may not include the third image group DB 17-3. Further, the registration unit 13 may perform processing of registering, in the second image group DB 17-2, an acquired image whose similarity to a first image is equal to or less than the first reference value, as a second image, and may not perform processing of registering, in the third image group DB 17-3, an acquired image whose similarity to a first image is equal to or more than the second reference value, as a third image. In this case, an image indicating a normal state is accumulated by processing by the registration unit 13.
Next, one example of a hardware configuration of the learning apparatus 10 is described. Each functional unit of the learning apparatus 10 is achieved by any combination of hardware and software mainly including a central processing unit (CPU) of any computer, a memory, a program loaded in a memory, a storage unit (capable of storing, in addition to a program stored in advance at a shipping stage of an apparatus, a program downloaded from a storage medium such as a compact disc (CD), a server on the Internet, and the like) such as a hard disk storing the program, and an interface for network connection. Further, it is understood by a person skilled in the art that there are various modification examples as a method and an apparatus for achieving the configuration.
The bus 5A is a data transmission path along which the processor 1A, the memory 2A, the peripheral circuit 4A, and the input/output interface 3A mutually transmit and receive data. The processor 1A is, for example, an arithmetic processing apparatus such as a CPU and a graphics processing unit (GPU). The memory 2A is, for example, a memory such as a random access memory (RAM) and a read only memory (ROM). The input/output interface 3A includes an interface for acquiring information from an input apparatus, an external apparatus, an external server, an external sensor, a camera, and the like, an interface for outputting information to an output apparatus, an external apparatus, an external server, and the like, and the like. The input apparatus is, for example, a keyboard, a mouse, a microphone, a physical button, a touch panel, and the like. The output apparatus is, for example, a display, a speaker, a printer, a mailer, and the like. The processor 1A can issue a command to each module, and perform an arithmetic operation, based on these arithmetic operation results.
Next, an advantageous effect of the learning apparatus 10 is described.
The learning apparatus 10 according to the present example embodiment generates an estimation model for discriminating between normal and abnormal by machine learning in which an image indicating a normal state and an image indicating an abnormal state are training images. In the estimation model, a regular state being observed during most of time is discriminated to be normal, and a state different from the regular state is discriminated to be abnormal.
In such a case, even when an abnormal state not being defined in advance has occurred, as far as the state is a state different from a normal state, the state can be discriminated as an abnormal state. Therefore, it becomes possible to detect an abnormal state without omission.
Further, in a case where an abnormal state is defined in advance, and an estimation model for detecting the abnormal state is generated, it is necessary to prepare a large number of training images indicating each abnormal state. However, it is not easy to prepare a training image indicating an abnormal state. In a case of the present example embodiment, as compared with a case where an estimation model for detecting an abnormal state defined in advance is generated, the number of “images indicating an abnormal state” required to be prepared decreases. Consequently, a load of a user is reduced.
Note that, in a case of the present example embodiment, a large number of “images indicating a normal state” are necessary. However, generally, since most of targets are in a “normal state”, it becomes possible to easily collect an “image indicating a normal state” from images in which such a target is photographed.
Further, in a case of the present example embodiment, it is possible to automatically accumulate an “image indicating a normal state”, based on a determination result of a similarity between a small number of “images indicating an abnormal state” (means that the number is smaller than the number of images indicating an abnormal state required for generating an estimation model for detecting an abnormal state defined in advance), which have been prepared in advance, and an image generated by a surveillance camera or the like. Therefore, a load of a user is reduced.
Further, in a case of the present example embodiment, it is possible to increase the number of “images indicating an abnormal state” by the second image registration processing S6. Since the number of “images indicating an abnormal state” can be increased as described above, estimation accuracy of an estimation model to be acquired improves.
Further, in a case of the present example embodiment, it is also possible to increase the number of “images indicating an abnormal state” by the first image registration processing S1. In this case, it is possible to increase the number of “images indicating an abnormal state” having higher reliability by setting the above-described second reference value to a sufficiently high value. Further, by increasing the number of “images indicating an abnormal state”, improvement of estimation accuracy of an estimation model to be acquired is expected.
Further, in a case of the present example embodiment, it is possible to classify and manage images indicating an abnormal state into “a first image being confirmed to indicate an abnormal state by a user, and having high reliability”, and “a third image determined to be similar to the first image by a predetermined level or more by a computer”. Further, it is possible to set only the first image as a reference target in the similarity computation S10 in
Further, in a case of the present example embodiment, it is possible to learn a plurality of estimation models concurrently. Therefore, it becomes possible to select and use an estimation model from which a more preferable result is acquired from among estimation models in an actual estimation scene (estimation by an estimation apparatus to be described in the following example embodiment).
In the first example embodiment, images indicating an abnormal state are classified and managed into “a first image being confirmed to indicate an abnormal state by a user, and having high reliability”, and “a third image determined to be similar to the first image by a predetermined level or more by a computer”. However, in the learning apparatus 10 according to the present example embodiment, such management is not performed. Specifically, “an image being confirmed to indicate an abnormal state by a user, and having high reliability”, and “an image determined to be similar to the image having high reliability by a predetermined level or more by a computer” are collectively managed as “a first image indicating an abnormal state”. The “first image” according to the present example embodiment is an image indicating an abnormal state, and conceptually including the first image and the third image described in the first example embodiment.
A registration unit 13 registers, in a first image group DB 17-1, an acquired image whose similarity to a first image registered in the first image group DB 17-1 is equal to or more than a second reference value, as a first image.
Other configurations of the learning apparatus 10 according to the present example embodiment are similar to those of the first example embodiment.
In the learning apparatus 10 according to the present example embodiment described above, an advantageous effect similar to that of the learning apparatus 10 according to the first example embodiment is achieved. Further, it is possible to efficiently collect an image indicating an abnormal state. Note that, reliability (reliability regarding indication of an abnormal state) of an “image being confirmed to indicate an abnormal state by a user, and having high reliability”, and an “image determined to be similar to the image having high reliability by a predetermined level or more by a computer” may be different from each other. Further, managing images having different reliability in a mixed manner may adversely affect learning accuracy, estimation accuracy, or the like. However, setting the above-described second reference value to a sufficiently high value enables to reduce such an inconvenience.
An estimation apparatus according to a present example embodiment discriminates a state (normal or abnormal) indicated by an image by using an estimation model generated by the learning apparatus 10 according to the first or second example embodiment.
Since the estimation apparatus according to the present example embodiment can collect training images of a sufficient number and having high accuracy by the unique method as described above, and use an estimation model generated by learning based on the training images, high estimation accuracy can be acquired.
As described above, while the example embodiments according to the present invention have been described with reference to the drawings, these example embodiments are an example of the present invention, and various configurations other than the above can also be adopted.
Further, in a plurality of flowcharts used in the above description, a plurality of processes (pieces of processing) are described in order, but an order of execution of processes to be executed in each example embodiment is not limited to the order of description. In each example embodiment, the illustrated order of processes can be changed within a range that does not adversely affect a content. Further, the above-described example embodiments can be combined, as far as contents do not conflict with each other.
A part or all of the above-described example embodiments may also be described as the following supplementary notes, but is not limited to the following.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2020/024793 | 6/24/2020 | WO |
