Target detection method based on foreground and background splitting and learning path feature distillation, storage medium and equipment

Abstract

The invention discloses a target detection method based on foreground and background splitting and learning path feature distillation, a storage medium and equipment, and the method comprises the steps: collecting different types of pictures, and carrying out the target frame and target type marking of each picture; respectively inputting the marked picture into a teacher model and a student model, carrying out feature distillation based on foreground and background splitting and a learning path, and guiding the training of the student model according to a feature distillation loss function between the teacher model and the student model until a total loss function is converged, thereby completing the training of the student model; and inputting the pictures collected in real time into the trained student model for reasoning to obtain the category and position information of the target. The knowledge of the teacher model is migrated to the student model, so that the target detection performance of the student model approaches to the teacher model through foreground and background splitting and learning path feature distillation on the premise of not increasing the parameter quantity of the student model, and the method has the characteristics of low resource consumption, high reasoning speed and suitability for mobile equipment.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a target detection method, storage medium and device based on foreground-background splitting and learning path feature distillation. Background Art

[0002] Object detection is a computer vision technology that can identify and locate specific objects in images or videos. This technology has important applications in many fields, including industry, medicine, transportation, etc. First of all, object detection plays a key role in the field of autonomous driving. Autonomous vehicles rely on object detection to identify other vehicles, pedestrians, traffic lights and signs on the road, which ensures the safe operation of vehicles in complex traffic environments. Secondly, object detection is also very important in security monitoring. In monitoring systems in public places or sensitive areas, object detection can help identify potential threats, such as suspicious people or drones, which helps prevent criminal activities and provide rapid response in emergencies.

[0003] Early object detection methods, such as the R-CNN series, used a region proposal method to generate possible candidate regions and then use CNN to classify each region. Although this method has high accuracy, it has a large computational cost and relatively slow processing speed.

[0004] The YOLO series changed this approach and used a single neural network for end-to-end detection. It takes the entire image as input and then directly outputs multiple bounding boxes and corresponding category predictions. This design enables YOLO to detect at an extremely fast speed while maintaining high accuracy. The emergence of YOLO has led to the development of target detection technology, and many subsequent target detection methods have adopted a similar architecture. Although the YOLO series of algorithms have been widely recognized in the industry and have achieved a good trade-off between detection accuracy and speed, it is still difficult to apply on many edge and mobile devices due to the limitations of network parameters: for example, from YOLOv3 to YOLOv8, the model structure has become increasingly complex, including convolutional layers, feature extraction layers, residual modules, cross-layer connections, etc. This structure increases the number of parameters and computational complexity of the model; for example, the number of parameters of YOLOv6 may be millions in small models, and may reach tens of millions or even higher in larger models, while mobile devices usually use low-power ARM processors, especially when processing deep learning models, their computing power is far inferior to high-performance GPUs or CPUs. Therefore, when a complex network like YOLO is inferred on a mobile device, it will take too long to infer, affecting the user experience. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a target detection method, storage medium and device based on foreground-background splitting and learning path feature distillation, which transfer the knowledge of the teacher model to the student model, so that the target detection performance of the student model is close to that of the teacher model through foreground-background splitting and learning path feature distillation without increasing the number of student model parameters, and has the characteristics of low resource consumption, fast inference speed and suitability for mobile devices.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions: A target detection method based on foreground-background segmentation and learning path feature distillation specifically includes the following steps: Step 1: Collect target detection images of different categories, mark each target detection image with a target frame and target category, and obtain a marked image; Step 2: Input the labeled images into the teacher model and the student model respectively, perform feature distillation based on the foreground-background splitting and the learning path, and guide the training of the student model according to the feature distillation loss function between the teacher model and the student model until the total loss function converges, thus completing the training of the student model. Step 3: Input the real-time collected target detection images into the trained student model for inference to obtain the target category and location information.

[0007] Furthermore, step 2 includes the following sub-steps: Step 2.1, input the labeled images into the teacher model and the student model respectively to perform learning path feature distillation, calculate the learning paths of the teacher model and the student model respectively, and establish the learning path feature distillation loss function according to the learning paths of the teacher model and the student model; Step 2.2, perform foreground-background split feature distillation on the marked image in the teacher model and the student model, output the teacher model feature map and the student model feature map, and establish the foreground-background split feature distillation loss function based on the teacher model feature map and the student model feature map; Step 2.3: Weight the learning path feature distillation loss function established in step 2.1 and the foreground-background split feature distillation loss function established in step 2.2 to obtain the total feature distillation loss function. Combine the total feature distillation loss function with the category prediction loss function and the target box prediction loss function to establish the total loss function for student model training. Step 2.4: Repeat steps 2.1-2.3 until the total loss function converges and the training of the student model is completed.

[0008] Furthermore, step 2.1 includes the following sub-steps: Step 2.1.1, input the marked images into the teacher model and the student model respectively, obtain the backbone network output feature map of the teacher model, the neck network output feature map of the teacher model, the backbone network output feature map of the student model, and the neck network output feature map of the student model; Step 2.1.2, calculate the learning path between the backbone network and the neck network of the teacher model according to the backbone network output feature map of the teacher model and the neck network output feature map of the teacher model, and calculate the learning path between the backbone network and the neck network of the student model according to the backbone network output feature map of the student model and the neck network output feature map of the student model; Step 2.1.3, establish a learning path feature distillation loss function based on the learning path between the backbone network to the neck network of the teacher model and the learning path between the backbone network to the neck network of the student model.

[0009] Furthermore, the learning path feature distillation loss function established in step 2.1.3 is L fsp Specifically: , in, N represents the total number of learning paths, N = X × Y ; X represents the total number of channels in the backbone network, x express X The index of Y represents the total number of channels in the neck network, y express Y The index of Represents the teacher model T The backbone network of x Channel to the neck network y Learning paths for each channel, , H and W Represent the height and width of the predicted target box respectively, h express H The index of w express W The index of Represents the teacher model backbone network Tb No. x The predicted target box pixels in the channel ( h, w ), Represents the teacher model neck network Tn No. y The predicted target box pixels in the channel ( h, w )’s output features; Represents the student modelS The backbone network of x Channel to the neck network y Learning paths for each channel, , Represents the backbone network of the student model Sb No. x The predicted target box pixels in the channel ( h, w ), Representation of the student model neck network Sn No. y The predicted target box pixels in the channel ( h, w )’s output features; Represents the backbone network x Channel to neck network y The learning path feature distillation weights of each channel.

[0010] Furthermore, step 2.2 includes the following sub-steps: Step 2.2.1. Generate a binary mask matrix to divide the foreground and background of the labeled image, where each element in the binary mask matrix is , g Represents the target box marked in the marked image, Indicates the pixel coordinates in the marked image; Step 2.2.2: According to the area occupied by different targets in the marked image, a proportional mask matrix is ​​introduced to balance the loss ratio between targets. Each element in the proportional mask matrix is , H r and W r Respectively represent the height and width of the marked target box, Indicates the total number of pixels in the background area of ​​the marked image. ; Step 2.2.3, both the teacher model and the student model perform foreground-background split feature distillation on the labeled image according to the binary mask matrix and the proportional mask matrix, and output the teacher model feature map and the student model feature map; Step 2.2.4: Establish the foreground-background split feature distillation loss function based on the difference between the teacher model feature map and the student model feature map L fea : , in, C represents the number of network channels in the teacher model or student model, k express C The index of Represents the teacher model kThe network channels correspond to the pixel coordinates of the marked image The output features, Represents the student model k The network channels correspond to the pixel coordinates of the marked image The output features, α A hyperparameter representing the loss ratio of the foreground area, β A hyperparameter representing the loss ratio of background area.

[0011] Furthermore, the total feature distillation loss function L KD for: , in, express L fsp The weight of express L fea The weight of .

[0012] Furthermore, the total loss function of the student model is specifically: , in, L cls represents the target category prediction loss function, L reg represents the target box prediction loss function, λ Indicates balance L reg The hyperparameter of the ratio, ε Indicates balance L KD The hyperparameter of the ratio.

[0013] Further, ε The specific calculation process is: , in, θ Represents the weight ratio of the total feature distillation loss, Indicates the current training round number of the student model, epochs Indicates the total number of rounds of student model training.

[0014] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the target detection method based on foreground-background segmentation and learning path feature distillation.

[0015] Furthermore, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the target detection method based on foreground-background segmentation and learning path feature distillation is implemented.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The target detection method based on foreground-background splitting and learning path feature distillation of the present invention performs learning path feature distillation on the teacher model and the student model respectively, thereby capturing the learning path feature distillation loss between the teacher model and the student model, so that the student model pays attention to the intermediate process of the teacher model imparting knowledge during the learning process, thereby helping the student model to understand the knowledge provided by the teacher model more quickly; at the same time, the feature graph output by the teacher model and the feature graph output by the student model are split foreground-background, respectively capturing the distillation loss of the foreground target features and the background features between the teacher model and the student model, so that the student model pays more attention to the details of the foreground target during the learning process. Based on this, without increasing the number of student model parameters, the present invention approaches the target detection performance of the student model to the teacher model, improves the reasoning accuracy of the student model, and thus realizes high-precision target detection of the student model. Therefore, the target detection method of the present invention is applied to the field of autonomous driving, which can improve the target detection accuracy and ensure safe driving of the vehicle; the target detection method of the present invention is applied to the field of security monitoring, which can accurately identify potential threatening targets and prevent crimes. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the target detection method based on foreground-background splitting and learning path feature distillation of the present invention; Figure 2 Schematic diagram of feature distillation based on foreground-background splitting and learning paths in the teacher model and the student model in the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings.

[0019] like Figure 1 This is a flow chart of the target detection method based on foreground-background splitting and learning path feature distillation of the present invention. The target detection method specifically includes the following steps: Step 1: Randomly extract target detection pictures of different categories from the video website, mark each picture with a target frame and target category, and obtain a marked picture.

[0020] Step 2: Input the labeled images into the teacher model and the student model respectively, perform feature distillation based on foreground-background splitting and learning path, and guide the training of the student model according to the feature distillation loss function between the teacher model and the student model until the total loss function converges and the training of the student model is completed. This method enables the student model to pay more attention to the details of the detection target and the intermediate process of the teacher model imparting knowledge during the learning process, thereby improving the reasoning speed and accuracy; Figure 2 , specifically including the following sub-steps: Step 2.1, the teacher model and the student model both include a backbone network, a neck network and a head network. When guiding the training of the student model, the existing teacher model only provides the final output of the teacher model to the student model for learning, but does not teach the student model how the teacher model obtains the intermediate process of the final output, which results in the student model being unable to quickly absorb the knowledge provided by the teacher model. Therefore, the present invention proposes to input the labeled images into the teacher model and the student model respectively for learning path feature distillation, calculate the learning paths of the teacher model and the student model respectively, and establish a learning path feature distillation loss function according to the learning paths of the teacher model and the student model. The learning path feature distillation loss function enables the student model to pay attention to the intermediate process of the teacher model imparting knowledge during the learning process, thereby helping the student model to understand the knowledge provided by the teacher model more quickly; specifically, the following sub-steps are included: Step 2.1.1, input the marked images into the teacher model and the student model respectively, obtain the backbone network output feature map of the teacher model, the neck network output feature map of the teacher model, the backbone network output feature map of the student model, and the neck network output feature map of the student model; Step 2.1.2, calculate the learning path between the backbone network and the neck network of the teacher model according to the backbone network output feature map of the teacher model and the neck network output feature map of the teacher model, and calculate the learning path between the backbone network and the neck network of the student model according to the backbone network output feature map of the student model and the neck network output feature map of the student model; Step 2.1.3: Establish a learning path feature distillation loss function based on the learning path between the backbone network and the neck network of the teacher model and the learning path between the backbone network and the neck network of the student model. L fsp : , in, N represents the total number of learning paths, N = X × Y ; X represents the total number of channels in the backbone network, x express X The index of Y represents the total number of channels in the neck network,y express Y The index of Represents the teacher model T The backbone network of x Channel to the neck network y Learning paths for each channel, , H and W Represent the height and width of the predicted target box respectively, h express H The index of w express W The index of Represents the teacher model backbone network Tb No. x The predicted target box pixels in the channel ( h, w ), Represents the teacher model neck network Tn No. y The predicted target box pixels in the channel ( h, w )’s output features; Represents the student model S The backbone network of x Channel to the neck network y Learning paths for each channel, , Represents the backbone network of the student model Sb No. x The predicted target box pixels in the channel ( h, w ), Representation of the student model neck network Sn No. y The predicted target box pixels in the channel ( h, w )’s output features; Represents the backbone network x Channel to neck network y The learning path feature distillation weights of the channels can be dynamically adjusted through softmax. and The greater the difference, The bigger.

[0021] Step 2.2, in the target detection task, if the probability distribution of the feature map of the corresponding layer of the teacher model and the student model is directly fitted according to the method of output position distillation, since the teacher model and the student model tend to focus on the background area with a large number of pixels in the picture, and the foreground target may only occupy a small area in the overall picture, the details of the foreground target are ignored, resulting in the student model being unable to learn the detailed representation of the foreground target by the teacher model. Therefore, the present invention performs foreground and background split feature distillation on the marked picture in the teacher model and the student model, outputs the teacher model feature map and the student model feature map, and establishes the foreground and background split feature distillation loss function based on the teacher model feature map and the student model feature map; specifically includes the following sub-steps: Step 2.2.1. Generate a binary mask matrix to divide the foreground and background of the labeled image, where each element in the binary mask matrix , g Represents the target box marked in the marked image, Indicates the pixel coordinates in the marked image; Step 2.2.2: After dividing the foreground and background, although the loss ratio between the foreground and the background is balanced, the problem of unbalanced loss ratio between different targets cannot be solved. Because the large target occupies more pixels, the calculated loss value is larger. Therefore, the student model will be more willing to optimize the large target and ignore the relatively small target. In order to overcome the above problem, the present invention adopts the strategy of inverse weighting of different target area ratios. According to the area occupied by different targets in the marked image, a proportional mask matrix is ​​introduced to balance the loss ratio between targets. Each element in the proportional mask matrix is , H r and W r Respectively represent the height and width of the marked target box, Indicates the total number of pixels in the background area of ​​the marked image. ; Step 2.2.3, both the teacher model and the student model perform foreground-background split feature distillation on the labeled image according to the binary mask matrix and the proportional mask matrix, and output the teacher model feature map and the student model feature map; Step 2.2.4: Establish the foreground-background split feature distillation loss function based on the difference between the teacher model feature map and the student model feature map L fea : , in, C represents the number of network channels in the teacher model or student model, k express C The index of Represents the teacher model k The network channels correspond to the pixel coordinates of the marked image The output features, Represents the student model k The network channels correspond to the pixel coordinates of the marked image The output features, α A hyperparameter representing the loss ratio of the foreground area, β A hyperparameter representing the loss ratio of background area.

[0022] Step 2.3: Weight the learning path feature distillation loss function established in step 2.1 and the foreground-background split feature distillation loss function established in step 2.2 to obtain the total feature distillation loss function. The total feature distillation loss function is combined with the category prediction loss function and the target box prediction loss function to establish the total loss function for student model training L total : , in, L cls represents the target category prediction loss function, L reg represents the target box prediction loss function, λ Indicates balance L reg The hyperparameter of the ratio, ε Indicates balance L KD In the early stage of student model training, the soft labels from the teacher model are easier to learn. However, as the student model training progresses, the performance of the student model usually exceeds that of the teacher model. At this time, the soft labels generated by the teacher model will have a certain impact on the training of the student model. The negative effect of this impact is relatively large. Based on this, ε Apply cosine weight decay to dynamically adjust the target prediction information from the image tags and the teacher model: , θ Indicates the weight ratio of the total feature distillation loss, usually θ The value is 0.01; Indicates the current training round number of the student model, epochs Indicates the total number of rounds of student model training.

[0023] Step 2.4: Repeat steps 2.1-2.3 until the total loss function converges and the training of the student model is completed.

[0024] Step 3: Input the real-time collected images into the trained student model for inference to obtain the category and location information of the target.

[0025] Compared with the YOLOv6N model, the target detection method based on foreground and background splitting and learning path feature distillation of the present invention has an increase of 3.1% in mAP@.5:.95 at the same reasoning speed; at the same time, without increasing the number of student model parameters, the target detection method based on foreground and background splitting and learning path feature distillation of the present invention has a faster reasoning speed than the YOLOv6N model, so the target detection method of the present invention has a higher reasoning speed and reasoning accuracy. Therefore, applying the target detection method of the present invention to the field of autonomous driving can improve the target detection accuracy and ensure safe driving of the vehicle; applying the target detection method of the present invention to the field of security monitoring can accurately identify potential threatening targets and prevent crimes.

[0026] In a technical solution of the present invention, a computer-readable storage medium is also provided, storing a computer program, wherein the computer program enables a computer to execute the target detection method based on foreground-background segmentation and learning path feature distillation.

[0027] In a technical solution of the present invention, an electronic device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the target detection method based on foreground-background segmentation and learning path feature distillation is implemented.

[0028] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0029] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0030] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A target detection method based on foreground-background splitting and learning path feature distillation, characterized in that: The specific steps include: Step 1: Collect target detection images of different categories, mark each target detection image with a target frame and target category, and obtain a marked image; Step 2: Input the labeled images into the teacher model and the student model respectively, perform feature distillation based on the foreground-background splitting and the learning path, and guide the training of the student model according to the feature distillation loss function between the teacher model and the student model until the total loss function converges, thus completing the training of the student model. Step 3: Input the real-time collected target detection images into the trained student model for inference to obtain the target category and location information.

2. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 1, characterized in that: Step 2 includes the following sub-steps: Step 2.1, input the labeled images into the teacher model and the student model respectively to perform learning path feature distillation, calculate the learning paths of the teacher model and the student model respectively, and establish the learning path feature distillation loss function according to the learning paths of the teacher model and the student model; Step 2.2, perform foreground-background split feature distillation on the marked image in the teacher model and the student model, output the teacher model feature map and the student model feature map, and establish the foreground-background split feature distillation loss function based on the teacher model feature map and the student model feature map; Step 2.3: Weight the learning path feature distillation loss function established in step 2.1 and the foreground-background split feature distillation loss function established in step 2.2 to obtain the total feature distillation loss function. Combine the total feature distillation loss function with the category prediction loss function and the target box prediction loss function to establish the total loss function for student model training. Step 2.4: Repeat steps 2.1-2.3 until the total loss function converges and the training of the student model is completed.

3. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 2, characterized in that: Step 2.1 includes the following sub-steps: Step 2.1.1, input the marked images into the teacher model and the student model respectively, obtain the backbone network output feature map of the teacher model, the neck network output feature map of the teacher model, the backbone network output feature map of the student model, and the neck network output feature map of the student model; Step 2.1.2, calculate the learning path between the backbone network and the neck network of the teacher model according to the backbone network output feature map of the teacher model and the neck network output feature map of the teacher model, and calculate the learning path between the backbone network and the neck network of the student model according to the backbone network output feature map of the student model and the neck network output feature map of the student model; Step 2.1.3, establish a learning path feature distillation loss function based on the learning path between the backbone network to the neck network of the teacher model and the learning path between the backbone network to the neck network of the student model.

4. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 3, characterized in that: The learning path feature distillation loss function established in step 2.1.3 L fsp Specifically: ， in, N represents the total number of learning paths, N=X × Y ; X represents the total number of channels in the backbone network, x express X The index of Y represents the total number of channels in the neck network, y express Y The index of Represents the teacher model T The backbone network of x Channel to the neck network y Learning paths for each channel, , H and W Represent the height and width of the predicted target box respectively, h express H The index of w express W The index of Represents the teacher model backbone network Tb No. x The predicted target box pixels in the channel ( h,w ), Represents the teacher model neck network Tn No. y The predicted target box pixels in the channel ( h,w )’s output features; Represents the student model S The backbone network of x Channel to the neck network y Learning paths for each channel, , Represents the backbone network of the student model Sb No. x The predicted target box pixels in the channel ( h,w ), Representation of the student model neck network Sn No. y The predicted target box pixels in the channel ( h,w )’s output features; Represents the backbone network x Channel to neck network y The learning path feature distillation weights of each channel.

5. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 4, characterized in that: Step 2.2 includes the following sub-steps: Step 2.2.

1. Generate a binary mask matrix to divide the foreground and background of the labeled image, where each element in the binary mask matrix is , g Represents the target box marked in the marked image, Indicates the pixel coordinates in the marked image; Step 2.2.2: According to the area occupied by different targets in the marked image, a proportional mask matrix is ​​introduced to balance the loss ratio between targets. Each element in the proportional mask matrix is , H r and W r Respectively represent the height and width of the marked target box, Indicates the total number of pixels in the background area of ​​the marked image. ; Step 2.2.3, both the teacher model and the student model perform foreground-background split feature distillation on the labeled image according to the binary mask matrix and the proportional mask matrix, and output the teacher model feature map and the student model feature map; Step 2.2.4: Establish the foreground-background split feature distillation loss function based on the difference between the teacher model feature map and the student model feature map L fea : ， in, C represents the number of network channels in the teacher model or student model, k express C The index of Represents the teacher model k The network channels correspond to the pixel coordinates of the labeled image The output features, Represents the student model k The network channels correspond to the pixel coordinates of the labeled image The output features, α A hyperparameter representing the loss ratio of the foreground area, β A hyperparameter representing the loss ratio of background area.

6. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 5, characterized in that: Total feature distillation loss function L KD for: ， in, express L fsp The weight of express L fea The weight of .

7. The object detection method based on foreground-background segmentation and learning path feature distillation according to claim 6, characterized in that: The total loss function of the student model is specifically: ， in, L cls represents the target category prediction loss function, L reg represents the target box prediction loss function, λ Indicates balance L reg The hyperparameter of the proportion, ε Indicates balance L KD The hyperparameter of the ratio.

8. The target detection method based on foreground-background segmentation and learning path feature distillation according to claim 7, characterized in that: ε The specific calculation process is: ， in, θ Represents the weight ratio of the total feature distillation loss, Indicates the current training round number of the student model, epochs Indicates the total number of rounds of student model training.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the target detection method based on foreground-background segmentation and learning path feature distillation as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the target detection method based on foreground-background splitting and learning path feature distillation as described in any one of claims 1 to 8 is implemented.

Images