Deep learning based image recognition method and system

Abstract

The application provides a kind of based on deep learning's image recognition method and system, belong to image data processing field.Based on deep learning's image recognition method, comprising: obtaining first image, second image and third image;With pre-trained occlusion identification model, according to first image, second image and third image, generate the pixel occlusion classification graph of first image;Obtain the depth map of first image;According to first image, the pixel occlusion classification graph of first image and the depth map of first image, identify the target object of first image.In the embodiment of the application, one or more technical solutions can improve the accuracy of identifying target objects.

Description

Technical Field

[0001] This invention relates to the field of image data processing, and more specifically to an image recognition method and system based on deep learning. Background Technology

[0002] Image recognition technology is an important field of artificial intelligence. It refers to the technology of object recognition in images to identify targets and objects of various patterns. With the development of artificial intelligence, image recognition technology is constantly being updated.

[0003] Chinese patent application CN110084244A discloses an image-based object recognition method. This method is applied to identify objects within the cavity of a smart device. An image acquisition device is installed within the smart device cavity to acquire images of the objects to be identified. The method includes the following steps: acquiring images of the objects to be identified within the smart device cavity acquired by the image acquisition device; inputting the images into a deep convolutional neural network model; and using the deep convolutional neural network model to identify the image and obtain the type, weight, quantity, and / or location of the objects. Because the type, weight, quantity, and / or location of objects are identified by combining the image acquisition device and the deep convolutional neural network model, the use of a weight sensor is avoided, thus reducing costs.

[0004] Chinese Patent Publication No. CN114202771A discloses an image recognition system and method for recognizing images of pigs. The image recognition system includes: one or more processors; a deep high-resolution network model; and a computer-readable storage medium storing program instructions for implementing the deep high-resolution network model. When the program instructions are executed by the one or more processors, the deep high-resolution network model performs the following: receiving image data related to the pig image; and performing operations on the image data to identify the pig target and key points within the pig target. Using the solution disclosed herein, the pig target and its key points can be obtained simultaneously, thereby reducing the consumption of computational resources.

[0005] Chinese patent application CN113191371A discloses an intelligent image recognition processing method and terminal. The method involves acquiring an image to be recognized, performing format conversion and preprocessing on the image to obtain a mean depth map; extracting and matching features from the mean depth map and a reference image; determining whether the mean depth map has relative rotation and adjusting it based on the extracted feature points; matching the reference image with the corner points of the adjusted mean depth map and overlapping feature points; mapping the overlapped mean depth map and the reference image to a reference coordinate system; and calculating the corresponding similarity value based on the obtained coordinate values ​​to complete the recognition process and improve the image recognition processing effect.

[0006] In practical applications, such as when an AI robot is identifying a target object in an image, if the target object is partially occluded, the aforementioned image recognition methods struggle to accurately identify the object, resulting in a low accuracy rate. While the invention patent with publication number CN114495229A considers occlusion, it merely adjusts the model parameters of the image recognition model using preset occlusion labels, which still fails to improve the accuracy of target object recognition. Summary of the Invention

[0007] To address at least one technical problem in the prior art, the present invention provides an image recognition method and system based on deep learning, thereby improving the accuracy of target object recognition.

[0008] According to a first aspect of the present invention, a deep learning-based image recognition method is provided for recognizing target objects in an image, the method comprising:

[0009] Acquire a first image, a second image, and a third image. The first image and the second image are images captured by the first camera and the second camera at the same time, respectively. The third image is the frame preceding the first image. The first camera and the second camera are separated by a preset distance and have the same shooting direction.

[0010] Using a pre-trained occlusion recognition model, a pixel occlusion classification map of the first image is generated based on the first image, the second image, and the third image. The pixel occlusion classification map is used to indicate whether each pixel belongs to the occluded area of ​​the target object. The occlusion recognition model is trained based on a first sample training set, which includes multiple first samples. The first sample includes a first image to be trained, a second image to be trained, a third image to be trained, and a pixel occlusion classification map of the first image to be trained. The second image to be trained of the first image to be trained is an image captured by a third camera and a fourth camera at the same time. The third image to be trained is the previous frame of the first image to be trained.

[0011] Obtain the depth map of the first image;

[0012] The target object in the first image is identified based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image.

[0013] Optionally, the occluded region identification model includes:

[0014] A first feature extraction layer is used to extract a first feature map from the first image;

[0015] The second feature extraction layer is used to extract the second feature map of the second image;

[0016] The third feature extraction layer is used to extract the third feature map of the third image;

[0017] The first fusion layer is used to fuse the first feature map, the second image map, and the third feature map to obtain the fourth feature map;

[0018] The first convolutional and deconvolutional layer is used to generate a pixel occlusion classification map of the first image based on the fourth feature map.

[0019] Optionally, obtaining the depth map of the first image includes: generating a depth map of the first image based on the first image and the second image.

[0020] Optionally, generating a depth map of the first image based on the first image and the second image includes:

[0021] A depth map of the first image is generated based on the first image and the second image, using a pre-trained depth recognition model.

[0022] Optionally, the method further includes:

[0023] Based on the first image, the depth map, and the pose change between the first image and the second image, a viewpoint is synthesized to form an image to be compared.

[0024] If the similarity between the image to be compared synthesized from the viewpoint and the first image is lower than a preset threshold, a warning message is generated.

[0025] Optionally, obtaining the depth map of the first image includes:

[0026] The depth map of the first image is obtained by scanning with the radar scanning module at the first camera, wherein the scanning area of ​​the radar scanning module includes the shooting area of ​​the first camera.

[0027] Optionally, identifying the target object in the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image includes:

[0028] Using a pre-trained image recognition model, the target object in the first image is identified based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image. The image recognition network is trained based on a second sample set, which includes multiple second samples. Each second sample includes a sample image, the pixel occlusion classification map of the sample image, the depth map of the sample image, and the target object label of the sample image.

[0029] Optionally, the image recognition model includes:

[0030] A depth variation feature extraction layer is used to extract the depth variation feature map of the first image based on the depth map of the first image;

[0031] The second fusion layer is used to fuse the first image, the depth variation feature map of the first image, and the pixel occlusion classification map of the first image to obtain a fused feature map.

[0032] The second convolutional and deconvolutional layers are used to identify and determine the target object in the first image based on the fused feature map.

[0033] According to a second aspect of the present invention, a deep learning-based image recognition system is provided for recognizing target objects in an image, the system comprising:

[0034] The acquisition module is used to acquire a first image, a second image, and a third image. The first image and the second image are images captured by a first camera and a second camera at the same time, respectively. The third image is the previous frame of the first image. The first camera and the second camera are separated by a preset distance and have the same shooting direction.

[0035] The first generation module is used to generate a pixel occlusion classification map of the first image based on the first image, the second image, and the third image using a pre-trained occlusion recognition model. The pixel occlusion classification map is used to indicate whether each pixel belongs to the occluded area of ​​the target object. The occlusion recognition model is trained based on a first sample training set, which includes multiple first samples. The first sample includes a first image to be trained, a second image to be trained, a third image to be trained, and a pixel occlusion classification map of the first image to be trained. The second image to be trained of the first image to be trained is an image captured by a third camera and a fourth camera at the same time. The third image to be trained is the previous frame of the first image to be trained.

[0036] The second generation module is used to generate a depth map of the first image based on the first image and the second image;

[0037] The recognition module is used to identify the target object in the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image.

[0038] Optionally, the occluded region identification model includes:

[0039] A first feature extraction layer is used to extract a first feature map from the first image;

[0040] The second feature extraction layer is used to extract the second feature map of the second image;

[0041] The third feature extraction layer is used to extract the third feature map of the third image;

[0042] The first fusion layer is used to fuse the first feature map, the second image map, and the third feature map to obtain the fourth feature map;

[0043] The first convolutional and deconvolutional layer is used to generate a pixel occlusion classification map of the first image based on the fourth feature map.

[0044] One or more technical solutions provided in the embodiments of this application utilize a first image, a pixel occlusion classification map of the first image, and a depth map of the first image to identify the target object in the first image. This can better take into account the occluded information, resulting in higher accuracy when identifying the target object. Attached Figure Description

[0045] The accompanying drawings illustrate exemplary embodiments of the invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification.

[0046] Figure 1 A schematic diagram of an implementation environment in which the various methods described herein may be implemented according to exemplary embodiments of the present invention is shown;

[0047] Figure 2 A flowchart of a deep learning-based image recognition method according to an exemplary embodiment of the present invention is shown;

[0048] Figure 3 A schematic block diagram of an occluded region identification model according to an exemplary embodiment of the present invention is shown;

[0049] Figure 4 A schematic block diagram of an image recognition model according to an exemplary embodiment of the present invention is shown;

[0050] Figure 5 A schematic block diagram of a deep learning-based image recognition system according to an exemplary embodiment of the present invention is shown. Detailed Implementation

[0051] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0052] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0053] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". It should be noted that the concepts of "first", "second", etc., mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of the functions performed by these devices, modules or units or their interdependencies.

[0054] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0055] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0056] The deep learning-based image recognition method of this invention can be applied to Figure 1 The implementation environment diagram is shown below. (See also...) Figure 1 The implementation environment includes two cameras 101 and 102, a client 103, and a server 104. The cameras, including spherical and cylindrical types, are used to continuously perform image capture operations. Optionally, after cameras 101 and 102 capture images, they can send the captured images to the terminal 103 or the server 104 for image recognition. The server can be used to train relevant models.

[0057] The present invention will now be described with reference to the accompanying drawings.

[0058] See Figure 2 Deep learning-based image recognition methods are used to identify target objects in images. These methods include:

[0059] S201, acquire a first image, a second image, and a third image. The first image and the second image are images captured by the first camera and the second camera at the same time, respectively. The third image is the previous frame of the first image. The first camera and the second camera are separated by a preset distance and have the same shooting direction.

[0060] The first and third images can be captured by the first camera, and the second image can be captured by the second camera. The camera can be a spherical, cylindrical, or other type. The preset distance can be set according to actual needs.

[0061] For example, the deep learning-based image recognition method of this embodiment can be applied to fields such as intelligent robots and autonomous driving to identify specific target objects, such as people or cars. When applied to intelligent robots, the first and second cameras can be two cameras positioned at the robot's eye level, and the robot's processing system can execute the method of this embodiment. When applied to autonomous driving, the first and second cameras can be two cameras positioned at the front of the vehicle, such as two cameras positioned below the two headlights, and the onboard processing system can execute the method of this embodiment.

[0062] S202, using a pre-trained occlusion recognition model, a pixel occlusion classification map of the first image is generated based on the first image, the second image, and the third image. The pixel occlusion classification map is used to indicate whether each pixel belongs to the occluded area of ​​the target object. The occlusion recognition model is trained based on a first sample training set, which includes multiple first samples. The first sample includes a first training image, a second training image, a third training image, and a pixel occlusion classification map of the first training image. The second training images of the first training image are images captured by the third camera and the fourth camera at the same time, respectively. The third training image is the previous frame of the first training image.

[0063] In this embodiment, the pixel occlusion classification map of the first image is used to indicate whether each pixel in the first image belongs to the occluded area of ​​the target object or not. It is understood that the method of this embodiment is applied to fields such as intelligent robots and autonomous driving. Robots, cars, and the target object being identified are often in motion. Furthermore, since the third image is the frame preceding the first image, the corresponding first camera and the corresponding subject may move when the first and third images are captured, resulting in slight changes in the occlusion situation. Therefore, the occlusion situations in the first and third images will differ. Similarly, since the first and second images are captured simultaneously by a first camera and a second camera at a preset distance, there are also slight differences in the occlusion situations in the first and second images. Therefore, when the occlusion recognition model generates the pixel occlusion classification map of the first image based on the first, second, and third images, a relatively accurate pixel occlusion classification map can be obtained.

[0064] As an alternative implementation, see Figure 3 The occlusion region recognition model includes: a first feature extraction layer 301 for extracting a first feature map from a first image; a second feature extraction layer 302 for extracting a second feature map from a second image; a third feature extraction layer 303 for extracting a third feature map from a third image; a first fusion layer 304 for fusing the first feature map, the second image map, and the third feature map to obtain a fourth feature map; and a first convolutional and deconvolutional layer 305 for generating a pixel occlusion classification map of the first image based on the fourth feature map. The first, second, and third feature extraction layers can each be composed of multiple convolutional blocks or multiple residual blocks cascaded together. The first fusion layer fuses the first feature map, the second image map, and the third feature map to obtain the fourth feature map. For example, feature fusion can be achieved by adding feature maps or increasing the number of channels. Increasing the number of channels involves directly concatenating the original features, allowing the network to learn how to fuse the features. This process does not result in information loss, although it is computationally intensive, it yields better results.

[0065] S203, Obtain the depth map of the first image.

[0066] As one implementation for obtaining the depth map of the first image, a depth map of the first image can be generated based on the first image and the second image using a pre-trained depth recognition model. The depth recognition model can be trained using images taken by two cameras spaced a first distance apart and their corresponding depth map labels as training samples. The depth recognition model in this implementation can be any existing depth recognition model; no limitation is made here.

[0067] To prevent inaccurate depth map generation, after step S203, a viewpoint-synthesized image to be compared is generated based on the first image, the depth map, and the pose change between the first and second images. If the similarity between the viewpoint-synthesized image to be compared and the first image is lower than a preset threshold, a warning message is generated.

[0068] As one implementation of obtaining the depth map of the first image, the depth map of the first image can be obtained by scanning with a radar scanning module at the first camera, wherein the scanning area of ​​the radar scanning module includes the shooting area of ​​the first camera.

[0069] S204, Identify the target object in the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image.

[0070] The pixel occlusion classification map of the first image reflects the information about the occlusion of the target object, and the depth map of the first image helps to determine the occlusion information. Therefore, by using the first image, its pixel occlusion classification map, and its depth map to identify the target object in the first image, the occlusion information can be better taken into account, resulting in higher accuracy when identifying the target object, especially when the target object is occluded by a large area. Compared with existing recognition methods, its accuracy is higher. When the target object in the first image is identified, a prompt message can be generated to remind the user.

[0071] As one implementation method for identifying the target object of the first image, a pre-trained image recognition model is used to identify the target object of the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image. The image recognition network is trained based on a second sample set, which includes multiple second samples. Each second sample includes a sample image, the pixel occlusion classification map of the sample image, the depth map of the sample image, and the target object label of the sample image.

[0072] In one implementation, the training process of the image recognition model is as follows: input the sample image, the previous frame image of the sample image, and the auxiliary image taken by a camera with the same shooting direction as the camera that took the sample image and at a preset distance into the occlusion region recognition model; input the sample image and the auxiliary image into the depth recognition model; input the sample image, the output of the occlusion region recognition model, and the output of the depth recognition model into the image recognition model; calculate the loss function based on the output of the image recognition model and the target object label of the sample image; and adjust the parameters of the occlusion region recognition model, the depth recognition model, and the image recognition model based on the loss function.

[0073] As one implementation of an image recognition model, see Figure 4The image recognition model includes: a depth variation feature extraction layer 401, used to extract a depth variation feature map of the first image based on the depth map of the first image; a second fusion layer 402, used to fuse the first image, the depth variation feature map of the first image, and the pixel occlusion classification map of the first image to obtain a fused feature map; and a second convolution and deconvolution layer 403, used to identify and determine the target object in the first image based on the fused feature map. The depth variation feature extraction layer can extract the depth variation feature map of the first image, specifically based on the sum of the squares of the differences between the depth values ​​of a pixel and the depth values ​​of its adjacent pixels. The adjacent pixels can be the eight pixels corresponding to the pixel in the top, bottom, left, right, top-left, top-right, bottom-left, and bottom-right positions. The second fusion layer, when fusing the first image, the depth variation feature map of the first image, and the pixel occlusion classification map of the first image, can achieve feature fusion by adding the feature maps or increasing the number of channels.

[0074] See Figure 5 The present invention also provides a deep learning-based image recognition system for identifying target objects in images, comprising:

[0075] The acquisition module 501 is used to acquire a first image, a second image and a third image. The first image and the second image are images captured by the first camera and the second camera at the same time, respectively. The third image is the previous frame of the first image. The first camera and the second camera are separated by a preset distance and have the same shooting direction.

[0076] The first generation module 502 is used to generate a pixel occlusion classification map of the first image based on the first image, the second image, and the third image using a pre-trained occlusion recognition model. The pixel occlusion classification map is used to indicate whether each pixel belongs to the occluded area of ​​the target object. The occlusion recognition model is trained based on a first sample training set, which includes multiple first samples. The first sample includes a first image to be trained, a second image to be trained, a third image to be trained, and a pixel occlusion classification map of the first image to be trained. The second image to be trained is an image captured by the third camera and the fourth camera at the same time, respectively. The third image to be trained is the previous frame of the first image to be trained.

[0077] The second generation module 503 is used to generate a depth map of the first image based on the first image and the second image;

[0078] The recognition module 504 is used to identify the target object in the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image.

[0079] In one embodiment, the occlusion region recognition model includes: a first feature extraction layer for extracting a first feature map of a first image; a second feature extraction layer for extracting a second feature map of a second image; a third feature extraction layer for extracting a third feature map of a third image; a first fusion layer for fusing the first feature map, the second image map, and the third feature map to obtain a fourth feature map; and a first convolution and deconvolution layer for generating a pixel occlusion classification map of the first image based on the fourth feature map.

[0080] In one embodiment, the recognition module, when recognizing a target object in a first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image, specifically uses a pre-trained image recognition model to recognize the target object in the first image based on the first image, the pixel occlusion classification map of the first image, and the depth map of the first image. The image recognition network is trained based on a second sample set, which includes multiple second samples. Each second sample includes a sample image, a pixel occlusion classification map of the sample image, a depth map of the sample image, and a target object label for the sample image.

[0081] In one embodiment, the image recognition model includes: a depth variation feature extraction layer for extracting a depth variation feature map of the first image based on a depth map of the first image; a second fusion layer for fusing the first image, the depth variation feature map of the first image, and a pixel occlusion classification map of the first image to obtain a fused feature map; and a second convolution and deconvolution layer for identifying and determining a target object in the first image based on the fused feature map.

[0082] An exemplary embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform a method according to an embodiment of the present invention.

[0083] The following is a block diagram describing an electronic device that can serve as a server or client of this disclosure, which is an example of hardware devices that can be applied to various aspects of this disclosure. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the disclosure described and / or claimed herein.

[0084] Electronic devices include a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0085] An exemplary embodiment of the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0086] The above are merely specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any changes or modifications made by those skilled in the art within the scope of the present invention are covered by the patent scope of the present invention.

Claims

1. A deep learning-based image recognition method for recognizing a target object in an image, characterized by, The method comprises: acquiring a first image, a second image and a third image, the first image and the second image being images captured by a first camera and a second camera at the same time respectively, the third image being a previous frame image of the first image, the first camera and the second camera being apart by a preset distance and having the same shooting direction; generating a pixel occlusion classification map of the first image by using a pre-trained occlusion recognition model according to the first image, the second image and the third image, the pixel occlusion classification map being used to indicate whether each pixel belongs to an occluded region of a target object, wherein the occlusion recognition model is trained according to a first sample training set, the first sample training set comprising a plurality of first samples, the first sample comprising a first training image, a second training image, a third training image and a pixel occlusion classification map of the first training image, the first training image and the second training image being images captured by a third camera and a fourth camera at the same time respectively, the third training image being a previous frame image of the first training image; acquiring a depth map of the first image; identifying a target object of the first image according to the first image, the pixel occlusion classification map of the first image and the depth map of the first image; the occluded region recognition model comprises: a first feature extraction layer used to extract a first feature map of the first image; a second feature extraction layer used to extract a second feature map of the second image; a third feature extraction layer used to extract a third feature map of the third image; a first fusion layer used to fuse the first feature map, the second feature map and the third feature map to obtain a fourth feature map; a first convolution and deconvolution layer used to generate the pixel occlusion classification map of the first image according to the fourth feature map; the acquiring of the depth map of the first image comprises: generating the depth map of the first image according to the first image and the second image; the generating of the depth map of the first image according to the first image and the second image comprises: generating the depth map of the first image according to the first image and the second image by using a pre-trained depth recognition model; the method further comprises: synthesizing a to-be-compared image according to the first image, the depth map and a pose change between the first image and the second image; generating a warning information when a similarity between the to-be-compared image synthesized by the viewpoint and the first image is lower than a preset threshold; the identifying of the target object of the first image according to the first image, the pixel occlusion classification map of the first image and the depth map of the first image comprises: identifying the target object of the first image according to the first image, the pixel occlusion classification map of the first image and the depth map of the first image by using a pre-trained image recognition model, wherein the image recognition network is trained according to a second sample set, the second sample set comprising a plurality of second samples, the second sample comprising a sample image, a pixel occlusion classification map of the sample image, a depth map of the sample image and a target object label of the sample image; The image recognition model comprises: a depth change feature extraction layer configured to extract a depth change feature map of the first image according to a depth map of the first image; a second fusion layer configured to fuse the first image, the depth change feature map of the first image and a pixel occlusion classification map of the first image to obtain a fused feature map; a second convolution and deconvolution layer configured to recognize a target object of the first image according to the fused feature map.

2. The deep learning-based image recognition method of claim 1, wherein, The depth map of the first image comprises: scanning the depth map of the first image by using a radar scanning module at the first camera, wherein a scanning area of the radar scanning module contains a shooting area of the first camera.

3. A deep learning-based image recognition system for recognizing a target object in an image, characterized by, The system comprises: an acquisition module configured to acquire a first image, a second image and a third image, the first image and the second image being images shot by a first camera and a second camera at the same time, the third image being a previous frame image of the first image, the first camera and the second camera being apart from each other by a preset distance and having the same shooting direction; a first generation module configured to generate a pixel occlusion classification map of the first image by using a pre-trained occlusion recognition model according to the first image, the second image and the third image, the pixel occlusion classification map being configured to indicate whether each pixel belongs to an occluded area of a target object, wherein the occlusion recognition model is trained according to a first sample training set, the first sample training set comprising a plurality of first samples, the first sample comprising a first training image, a second training image, a third training image and a pixel occlusion classification map of the first training image, the first training image and the second training image being images shot by a third camera and a fourth camera at the same time, the third training image being a previous frame image of the first training image; a second generation module configured to generate a depth map of the first image according to the first image and the second image; an identification module configured to identify a target object of the first image according to the first image, the pixel occlusion classification map of the first image and the depth map of the first image; The occluded area recognition model comprises: a first feature extraction layer configured to extract a first feature map of the first image; a second feature extraction layer configured to extract a second feature map of the second image; a third feature extraction layer configured to extract a third feature map of the third image; a first fusion layer configured to fuse the first feature map, the second feature map and the third feature map to obtain a fourth feature map; a first convolution and deconvolution layer configured to generate the pixel occlusion classification map of the first image according to the fourth feature map; The depth map of the first image comprises: generating the depth map of the first image according to the first image and the second image; The depth map of the first image comprises: generating the depth map of the first image according to the first image and the second image according to a pre-trained depth recognition model; The system is further configured to: According to the first image, the depth map and the pose change between the first image and the second image, a view-synthesized image to be compared is synthesized; If the similarity between the view-synthesized image to be compared and the first image is lower than a preset threshold, a warning information is generated; The recognition module is configured to recognize a target object in the first image according to the first image, a pixel occlusion classification map of the first image and a depth map of the first image, and specifically configured to: The recognition module is configured to recognize a target object in the first image according to the first image, a pixel occlusion classification map of the first image and a depth map of the first image, and specifically configured to: The image recognition model includes: The depth change feature extraction layer is configured to extract a depth change feature map of the first image according to the depth map of the first image; The second fusion layer is configured to fuse the first image, the depth change feature map of the first image and the pixel occlusion classification map of the first image to obtain a fusion feature map; The second convolution and deconvolution layer is configured to recognize and determine the target object in the first image according to the fusion feature map.

Images