A work robot screen cabinet target detection method and system

Abstract

The application discloses a work robot screen cabinet target detection method and system, comprising: preprocessing a screen cabinet image; constructing a bidirectional network, which is trained based on a loss function comprising a screen cabinet target shape loss and a screen cabinet target position loss; using a first branch to sequentially extract feature maps from the preprocessed screen cabinet image, using a second branch to perform three-layer convolution operation on the preprocessed screen cabinet image with a stride of two to obtain 1 / 8 feature maps; performing first feature fusion on the 1 / 8 feature maps of the first branch and the 1 / 8 feature maps of the second branch, performing second feature fusion on the obtained fusion 1 / 8 feature maps and 1 / 32 feature maps to obtain a to-be-recognized feature map, and obtaining a detection result of the screen cabinet target according to the to-be-recognized feature map. The method avoids the interference of image artifacts or noise, reduces target detection error, has strong anti-interference ability and high recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation and target detection, and particularly relates to a work robot screen cabinet target detection method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Image segmentation is a technology of dividing the whole image into several independent local regions according to imaging characteristics such as gray distribution and tissue contrast. This technology can be used for understanding and analyzing the image of the work robot work area, detecting and positioning the work target, and measuring and evaluating the morphological characteristics of the target. There are a large number of image segmentation algorithms based on this. These algorithms can be roughly divided into unsupervised and supervised segmentation algorithms according to different image evaluation strategies.

[0004] Unsupervised segmentation algorithms usually distinguish the target region from the irrelevant background according to the inherent imaging characteristics of the image (such as gray distribution or tissue contrast), so as to accurately detect and extract the region of interest. This kind of algorithm usually has the characteristics of simple operation and less time-consuming, and can effectively process high-quality images.

[0005] However, this method is easily affected by various artifacts or noises in the image, so it is difficult to accurately detect the target of interest from images with serious imaging artifacts, noise or weak tissue contrast.

[0006] In addition, this kind of algorithm often contains more operation parameters that cannot be adaptively adjusted, which makes it difficult to be used in large-scale engineering applications. Generally, the parameters are assigned by experience, which limits the segmentation performance of the algorithm.

[0007] Supervised segmentation algorithms generally combine the inherent characteristics of the image and the manually labeled or selected information to distinguish and extract the target region from the irrelevant background. This kind of algorithm has lower influence of artifacts or noises on image segmentation due to the manual labeling of feature information, and obtains better segmentation performance than unsupervised algorithms. Among these segmentation algorithms, based on deep learning, end-to-end image segmentation can be automatically performed and high precision can be obtained. U-Net is the most classic convolutional neural network based on deep learning and is used for processing various images.

[0008] However, although U-Net network has good segmentation performance, it is difficult to process the boundary region of the target and has a large boundary detection error, because the network uses image downsampling operation multiple times, which greatly reduces the image resolution, resulting in the blurring of the target boundary. In addition, because the network model has many parameters, the running speed is slow and it cannot run in real time on common processors. SUMMARY

[0009] In order to solve the above problems, the present application provides a work robot screen cabinet target detection method and system, which avoids the interference of image artifacts or noise, reduces target detection error, has strong anti-interference ability and high recognition accuracy.

[0010] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0011] In a first aspect, the present application provides a work robot screen cabinet target detection method, comprising:

[0012] After feature extraction of point-line-surface, color space and geometric gradient of the obtained screen cabinet image, the screen cabinet image is spliced to obtain a pre-processed screen cabinet image;

[0013] A bidirectional network including a first branch and a second branch is constructed, and the bidirectional network is trained based on a designed loss function, wherein the loss function includes a screen cabinet target shape loss and a screen cabinet target position loss;

[0014] The first branch is used to extract 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps from the pre-processed screen cabinet image in sequence, and the second branch is used to perform three-layer convolution operation with a stride of two on the pre-processed screen cabinet image to obtain a 1 / 8 feature map;

[0015] The 1 / 8 feature map of the first branch is fused with the 1 / 8 feature map of the second branch, and the obtained fused 1 / 8 feature map is fused with the 1 / 32 feature map to obtain a to-be-recognized feature map, and the detection result of the screen cabinet target is obtained according to the to-be-recognized feature map.

[0016] As an optional implementation, the screen cabinet target shape loss L s is:

[0017] L s =∑(Var[X h ]+Var[Y v ])

[0018] Wherein, Var[X h ] is the detection variance of all screen cabinet targets in the horizontal direction, var[Y v ] is the detection variance of all screen cabinet targets in the vertical direction, X h is the set of horizontal center coordinates of all screen cabinet targets, and Y v is the set of vertical center coordinates of all screen cabinet targets.

[0019] As an optional implementation, the screen cabinet target position loss function L γ:

[0020] L γ =‖c s -c b ‖ 2

[0021] wherein c s and c b are the area of the region of the segmentation result and the detection result respectively.

[0022] As an optional implementation, the process of the first feature fusion includes: after the 1 / 8 feature map of the first branch and the 1 / 8 feature map of the second branch respectively pass through a convolution layer, a normalization layer and a ReLU activation function, splicing is completed, and then the channel adjustment of a 1*1 convolution layer and the ReLU activation function are performed, a fused 1 / 8 feature map is obtained.

[0023] As an optional implementation, the process of the second feature fusion includes:

[0024] The fused 1 / 8 feature map and the 1 / 32 feature map are connected, and after sequentially passing through a 1x1 convolution, a batch normalization and a ReLU activation function, a first fused feature is obtained.

[0025] After the first fused feature sequentially passes through a global average pooling, a 1x1 convolution, a ReLU activation function, a 1*1 convolution and a sigmoid activation function, a second fused feature is obtained.

[0026] The first fused feature and the second fused feature are subjected to a mul function to obtain a third fused feature.

[0027] The third fused feature and the first fused feature are added to obtain a feature map to be recognized.

[0028] As an optional implementation, the second branch includes three convolution layers with a stride of 2, and each convolution layer is connected with a batch normalization layer and a ReLU activation function.

[0029] As an optional implementation, the first branch includes a 4-layer convolution lightweight model and a global average pooling layer arranged at the end of the model to perform 32 times down-sampling and provide a maximum receptive field through global information.

[0030] In a second aspect, the present application provides a work robot screen cabinet target detection system, comprising:

[0031] A preprocessing module is configured to perform feature extraction of point lines, color space and geometric gradient on the obtained screen cabinet image, splice the screen cabinet image to obtain a preprocessed screen cabinet image.

[0032] The bidirectional network construction module is configured to construct a bidirectional network comprising a first branch and a second branch, the bidirectional network being trained based on a designed loss function, the loss function comprising a screen cabinet target shape loss and a screen cabinet target position loss.

[0033] The bidirectional network feature extraction module is configured to sequentially extract 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps from the preprocessed screen cabinet image by using the first branch, and obtain a 1 / 8 feature map by performing a three-layer convolution operation with a stride of two on the preprocessed screen cabinet image by using the second branch.

[0034] The feature fusion and target detection module is configured to perform first feature fusion on the 1 / 8 feature map of the first branch and the 1 / 8 feature map of the second branch, perform second feature fusion on the obtained fused 1 / 8 feature map and the 1 / 32 feature map, obtain a to-be-recognized feature map, and obtain a detection result of the screen cabinet target according to the to-be-recognized feature map.

[0035] In a third aspect, the present application provides an electronic device, comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0036] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] The present application innovatively proposes a screen cabinet target detection method for a work robot in the power industry, constructs a bidirectional network model comprising a large receptive field content branch and a small receptive field detail branch to extract global information and detail information of a screen cabinet image, designs a loss function considering screen cabinet target shape loss and screen cabinet target position loss to make the target detection result more accurate, avoid the interference of artifacts or noise in the image, reduce the target detection error, have strong anti-interference ability and high recognition accuracy, and improve the intelligent recognition level and processing efficiency of the work robot.

[0039] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0040] The drawings accompanying the specification of the present application form a part thereof, serve to provide further understanding of the present application, and together with the description of the exemplary embodiments of the present application and their description serve to explain the present application, and do not constitute an improper limitation of the present application.

[0041] Figure 1 A work robot screen cabinet target detection method flow chart is provided for the embodiment 1 of the present application.

[0042] Figure 2 A first feature fusion flow schematic diagram is provided for the embodiment 1 of the present application. DETAILED DESCRIPTION

[0043] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0044] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0045] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0046] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0047] Embodiment 1

[0048] The present embodiment provides a work robot screen cabinet target detection method, as shown in Figure 1 The present embodiment provides a work robot screen cabinet target detection method, as shown in

[0049] After feature extraction of point-line-surface, color space and geometric gradient on the obtained screen cabinet image, the screen cabinet image is spliced to obtain a pre-processed screen cabinet image;

[0050] A bidirectional network including a first branch and a second branch is constructed, the bidirectional network is trained based on a designed loss function, and the loss function includes a screen cabinet target shape loss and a screen cabinet target position loss;

[0051] The first branch is used to extract 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps from the pre-processed screen cabinet image in sequence, and the second branch is used to perform three-layer convolution operation with a stride of two on the pre-processed screen cabinet image to obtain a 1 / 8 feature map;

[0052] The 1 / 8 feature map of the first branch is first feature-fused with the 1 / 8 feature map of the second branch, and the obtained fused 1 / 8 feature map is second feature-fused with the 1 / 32 feature map, to obtain a to-be-recognized feature map, and a detection result of the screen cabinet target is obtained according to the to-be-recognized feature map.

[0053] In the embodiment, in view of the features of the screen cabinet image in the power industry, a series of pretreatments are performed on the screen cabinet image by relying on traditional image feature extraction, and the fast addition of geometric features is enhanced in advance, so as to improve the model training speed and fitting capability.

[0054] Specifically, the screen cabinet image is sequentially subjected to Hough circle detection, LSD line detection, H channel detection, X / Y direction gradient detection and feature point detection, feature extraction of point-line-surface, color space and geometric gradient is completed, and then the extracted features are spliced to obtain combined features.

[0055] Among them, the circular information is extracted by Hough circle detection, the straight line segment information is extracted by LSD line detection, the H channel feature in the HSV color space is extracted by H channel detection, the gradient information in the horizontal and vertical directions of the image is extracted by X / Y direction gradient detection, and the image feature points are detected by feature point detection.

[0056] It can be understood that, in addition to the above-mentioned Hough circle detection, LSD line detection, H channel detection, X / Y direction gradient detection and feature point detection, other technical means capable of achieving the same function can also be used in the pretreatment process, and the present embodiment only gives one implementation scheme and does not make any limitation.

[0057] Through the above processing, the feature extraction of point-line-surface, color space and geometric gradient of the screen cabinet image is completed, and the combined features with a dimension of (1280, 1280, 6) are formed. Then, the combined features are spliced with the RGB original image (1280, 1280, 3) of the screen cabinet, and are scaled to 640*640 resolution, and after being divided by 255, the image normalization operation is completed, to obtain the pretreated screen cabinet image (640, 640, 9).

[0058] In the embodiment, a bidirectional network including a large receptive field content branch and a small receptive field detail branch is constructed to extract global information and detail information of the pretreated screen cabinet image respectively, and the extracted features are fused.

[0059] In the embodiment, the bidirectional network is trained based on a designed loss function; since the screen cabinet targets such as rotary switches, round buttons, air switches and the like have regular geometric shapes and clear boundaries, for example, the round button is a standard symmetrical circle and the rotary switch is a regular geometric symmetrical pattern. Therefore, in order to make the detection result closer to the actual true situation, the loss function including the screen cabinet target shape loss and the screen cabinet target position loss is designed.

[0060] Wherein, the screen cabinet target shape loss L s is:

[0061] L s = (Var[X h ]+Var[Y v ])

[0062]

[0063]

[0064]

[0065]

[0066] Wherein, Var[X h ] is the detection variance of all screen cabinet targets in the horizontal direction, Var[Y v ] is the detection variance of all screen cabinet targets in the vertical direction, X h is the set of horizontal center coordinates of all screen cabinet targets, xi hi ∈X h represents the horizontal center coordinate of the ith detection element, Y v is the set of vertical center coordinates of all screen cabinet targets.

[0067] Wherein, the screen cabinet target position loss is for; assuming that the time sequence feature input is X, there is a position encoding P providing time sequence information, and X+P is taken as the auto-encoding input. The position encoding is a matrix given according to a trigonometric function:

[0068]

[0069]

[0070] Wherein, i and j are matrix position information, and d is the length of the second dimension of the matrix. The designed position encoding can provide information on the absolute position and the relative position of the time sequence.

[0071] In order to further improve the detection accuracy, the loss constraint information is provided from the position encoding information, and the screen cabinet target position loss function L γ is designed:

[0072] L γ =‖c s -c b ‖ 2

[0073] wherein c s and c b are the area of the region of the segmentation result and the detection result respectively.

[0074] The final designed loss function is: L = L γ + L s .

[0075] In the embodiment, the large receptive field content branch is designed. Since the existing research direction for improving the receptive field is mainly the pyramid pooling module, the pyramid type hollow pooling and the like, these technical means often consume more computing resources, resulting in low model execution efficiency. Therefore, the content path, i.e., the large receptive field content branch, is designed to mainly extract the global information of the image and obtain a large enough receptive field.

[0076] The large receptive field content branch is a 4-layer convolution lightweight model, which realizes 32 times of down-sampling. The 4-layer convolution lightweight model and the global average pooling are fully utilized to provide a large receptive field. A global average pooling is added at the end of the lightweight model to provide a maximum receptive field through the global context information.

[0077] Specifically, the pre-processed screen cabinet image is subjected to different down-sampling multiples, and 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps are extracted in turn to obtain 1 / 4 feature maps, 1 / 8 feature maps, 1 / 16 feature maps and 1 / 32 feature maps with dimensions of (4, 256, 120, 120), (4, 512, 60, 60), (4, 1024, 30, 30) and (4, 2048, 15, 15) respectively.

[0078] In the embodiment, the small receptive field detail branch is designed. In the real-time semantic segmentation task, the existing method is difficult to balance the spatial resolution and the receptive field. Therefore, the detail path, i.e., the small receptive field detail branch, is designed to retain the spatial scale of the original input image and encode rich feature information.

[0079] Specifically, the pre-processed screen cabinet image is subjected to three-layer convolution operation with a stride of 2, and then subjected to batch normalization and ReLU activation function to obtain 1 / 8 feature maps.

[0080] In the embodiment, the 1 / 8 feature maps of the first branch and the 1 / 8 feature maps of the second branch are fused to further improve and enhance the feature information. For example, Figure 2As shown, the 1 / 8 feature map of the first branch and the 1 / 8 feature map of the second branch are respectively subjected to a convolution layer CONV, a normalization layer Norm and a ReLU activation function, spliced, and finally subjected to channel adjustment of a 1*1 convolution layer CONV, subjected to a ReLU activation function, and output a fused 1 / 8 feature map.

[0081] In this embodiment, the obtained fused 1 / 8 feature map and 1 / 32 feature map are further fused; since the low-level feature has higher resolution and contains more position and detail information, but has less convolution and lower semantics and more noise; and the high-level feature has stronger semantic information but lower resolution and poorer perception of details. Therefore, fusing features of different scales is an important means to improve performance.

[0082] Specifically,

[0083] (1) The fused 1 / 8 feature map and the 1 / 32 feature map are connected, sequentially subjected to 1x1 convolution, batch normalization and ReLU activation function, and a first fused feature with a size of (4, 1024, 60, 60) is obtained; the connected features are pooled into a feature vector by batch normalization to balance the feature scale;

[0084] (2) The first fused feature is sequentially subjected to global average pooling, 1x1 convolution, ReLU activation function, 1*1 convolution and sigmoid activation function, and a second fused feature with a size of (4, 1024, 1, 1) is obtained;

[0085] (3) The first fused feature and the second fused feature are subjected to a mul function, and a third fused feature with a size of (4, 1024, 60, 60) is obtained;

[0086] (4) Finally, the third fused feature and the first fused feature are added, and a to-be-recognized feature map with a size of (4, 1024, 60, 60) is obtained;

[0087] (5) Thus, target recognition is performed based on the to-be-recognized feature map, and since the loss function including the screen cabinet target shape loss and the screen cabinet target position loss is designed in this embodiment, the accuracy of the detection result is ensured.

[0088] Embodiment 2

[0089] The embodiment provides a work robot screen cabinet target detection system, comprising:

[0090] The preprocessing module is configured to perform feature extraction of point lines, color space and geometric gradient on the obtained screen cabinet image, splice the screen cabinet image, and obtain a preprocessed screen cabinet image;

[0091] The bidirectional network construction module is configured to construct a bidirectional network comprising a first branch and a second branch, the bidirectional network being trained based on a designed loss function, the loss function comprising a screen cabinet target shape loss and a screen cabinet target position loss.

[0092] The bidirectional network feature extraction module is configured to sequentially extract 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps from the preprocessed screen cabinet image by using the first branch, and obtain a 1 / 8 feature map by performing a three-layer convolution operation with a stride of two on the preprocessed screen cabinet image by using the second branch.

[0093] The feature fusion and target detection module is configured to perform first feature fusion on the 1 / 8 feature map of the first branch and the 1 / 8 feature map of the second branch, perform second feature fusion on the obtained fused 1 / 8 feature map and the 1 / 32 feature map, obtain a to-be-recognized feature map, and obtain a detection result of a screen cabinet target according to the to-be-recognized feature map.

[0094] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules can be executed in a computer system such as a set of computer executable instructions as part of a system.

[0095] In more embodiments, there are also provided:

[0096] An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method described in Embodiment 1 is completed. For brevity, it will not be repeated here.

[0097] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSPs, application-specific integrated circuits ASICs, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0098] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, a part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0099] A computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method described in Embodiment 1 is completed.

[0100] The method in the embodiment 1 can be directly embodied by a hardware processor or a combination of hardware and software modules in the processor. The software modules can be located in a storage medium in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or the like. The storage medium is located in the memory, and the processor reads information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, no further detailed description is given here.

[0101] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0102] Although the specific embodiments of the present application are described above in combination with the drawings, the description is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for target detection on a display cabinet of a work robot, characterized in that, include: After extracting features of points, lines, surfaces, color space, and geometric gradient from the acquired screen cabinet image, it is stitched with the screen cabinet image to obtain the preprocessed screen cabinet image. A bidirectional network is constructed, comprising a first branch (large receptive field content branch) and a second branch (small receptive field detail branch), to extract global and detail information from the preprocessed screen cabinet image, respectively. The bidirectional network is trained based on a designed loss function, which includes screen cabinet target shape loss and screen cabinet target position loss. The first branch extracts 1 / 4, 1 / 8, 1 / 16, and 1 / 32 feature maps sequentially from the preprocessed screen cabinet image. The second branch performs a three-layer convolution operation with a stride of 2 on the preprocessed screen cabinet image to obtain the 1 / 8 feature map. The first feature fusion is performed between 1 / 8 of the feature map from the first branch and 1 / 8 of the feature map from the second branch. The resulting fused 1 / 8 feature map is then fused with the 1 / 32 feature map to obtain the feature map to be identified. The detection result of the target on the screen cabinet is obtained based on the feature map to be identified.

2. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The target shape loss of the screen cabinet for: in, The detection variance in the horizontal direction for all screen cabinet targets. The detection variance in the vertical direction for all screen cabinet targets. Let this be the set of horizontal center coordinates of all screen cabinet targets. This is the set of vertical center coordinates for all screen cabinet targets.

3. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The target position loss function of the screen cabinet : in, and These represent the area of ​​the segmentation result and the area of ​​the detection result, respectively.

4. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The first feature fusion process includes: passing the 1 / 8 feature map of the first branch and the 1 / 8 feature map of the second branch through a convolutional layer, a normalization layer and a ReLU activation function respectively, concatenating them, and then passing them through a 1*1 convolutional layer for channel adjustment and a ReLU activation function to obtain a fused 1 / 8 feature map.

5. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The second feature fusion process includes: The first fused feature is obtained by connecting and fusing the 1 / 8 feature map and the 1 / 32 feature map, and then passing them through a 1x1 convolution, batch normalization and ReLU activation function in sequence; The first fused feature is sequentially processed through global average pooling, 1x1 convolution, ReLU activation function, 1*1 convolution, and sigmoid activation function to obtain the second fused feature; The third fusion feature is obtained by passing the first fusion feature and the second fusion feature through the mul function; The third fusion feature is added to the first fusion feature to obtain the feature map to be identified.

6. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The second branch consists of three convolutional layers with a stride of 2, each of which is connected to a batch normalization layer and a ReLU activation function.

7. The target detection method for a work robot display cabinet as described in claim 1, characterized in that, The first branch includes a 4-layer convolutional lightweight model and a global average pooling layer at the end of the model to perform 32x downsampling and provide the maximum receptive field through global information.

8. A target detection system for a work robot display cabinet, characterized in that, include: The preprocessing module is configured to extract features of points, lines, surfaces, color space, and geometric gradients from the acquired screen cabinet image, and then stitch it with the screen cabinet image to obtain the preprocessed screen cabinet image. The bidirectional network construction module is configured to construct a bidirectional network including a first branch (large receptive field content branch) and a second branch (small receptive field detail branch), which respectively extracts global information and detail information of the preprocessed screen cabinet image. The bidirectional network is trained based on a designed loss function, which includes screen cabinet target shape loss and screen cabinet target position loss. The bidirectional network feature extraction module is configured to use the first branch to extract 1 / 4, 1 / 8, 1 / 16 and 1 / 32 feature maps sequentially from the preprocessed screen cabinet image, and use the second branch to perform a three-layer convolution operation with a stride of 2 on the preprocessed screen cabinet image to obtain the 1 / 8 feature map. The feature fusion and target detection module is configured to perform a first feature fusion by combining 1 / 8 of the feature map from the first branch with 1 / 8 of the feature map from the second branch, and then perform a second feature fusion by combining the resulting fused 1 / 8 feature map with 1 / 32 of the feature map to obtain a feature map to be identified. Based on the feature map to be identified, the detection result of the target on the screen cabinet is obtained.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-7.

Images