A deep learning target detection system stress testing method and device

Generate multi-target test samples through adversarial attack technology, evaluate the timeliness of deep learning object detection systems under extreme conditions, solve the problem of difficulty in evaluating timeliness in the existing technology, improve the testing efficiency and accuracy, and ensure target detection System reliability and stability.

CN118885402BActive Publication Date: 2024-12-03COMP APPL RES INST CHINA ACAD OF ENG PHYSICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411369109.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-29
Publication Date
2024-12-03
Estimated Expiration
2044-09-29

AI Technical Summary

Technical Problem

The prior art is difficult to effectively evaluate the timeliness of deep learning object detection systems in extreme scenarios, especially on platforms with limited computing power such as drones, which lead to increased system latency and affect reliability and security.

Method used

Adversarial attack technology is used to generate multi-objective test samples, increase the number of candidate boxes through iterative optimization, design the test process to evaluate the performance of the algorithm, generate test videos containing different number of detection objects, and record the target detection time to evaluate the system's limiting conditions. Timeliness.

Benefits of technology

It realizes the rapid construction of diversified extreme scenarios under the premise of ensuring background consistency, overcomes the problem of high manpower and material consumption in traditional methods, effectively evaluates the timeliness of algorithms, ensures the reliability and stability of the target detection system, and is suitable for unmanned object detection systems such as machine and industrial control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118885402B_ABST
    Figure CN118885402B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of artificial intelligence security and provides a method and device for stress testing a deep learning target detection system. The main purpose of the present invention is to solve the problem of timeliness in evaluating the target detection system under extreme conditions. The main scheme includes generating and processing multi-target test samples for the original image, constraining the size of the candidate frame in the image, and obtaining an image with an increased number of targets; the image with the optimized number of targets is subjected to actual application processing of the test method to obtain test cases and extreme performance test results. This method is suitable for evaluating the extreme performance of target detection systems such as drones, industrial control, and scene monitoring, and helps to determine whether the hardware configuration meets the use requirements or reduce costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence security and provides a deep learning target detection system stress testing method. Background Art

[0002] In recent years, artificial intelligence technology has developed rapidly, especially in the field of computer vision. Artificial intelligence algorithms represented by deep learning have been rapidly iterating, and have reached or even surpassed the level of humans in areas such as image classification and target recognition. The tasks that current artificial intelligence models can handle are becoming more complex, and the situations they face are also more diverse. In extreme scenarios, artificial intelligence algorithms may lose functionality and degrade performance. However, extreme scenarios rarely occur in the real world. Therefore, constructing extreme inputs for artificial intelligence algorithms and testing the functions and performance of artificial intelligence algorithms in special scenarios have positive significance for promoting the application of artificial intelligence technology.

[0003] For deep learning-based target recognition algorithms, the scene uncertainty is extremely large, and the number of targets in the corresponding scenes also varies greatly. Currently, non-maximum suppression (NMS) is widely used in end-to-end target recognition deep learning algorithms (such as YOLO) to filter overlapping bounding boxes output by the model. The NMS algorithm needs to iterate all candidate bounding boxes until they are all processed. Therefore, as the number of targets increases, the number of NMS algorithm inputs will also increase accordingly, resulting in increased system latency. For devices that currently deploy target recognition algorithms, especially platforms with limited computing power such as drones, the increased latency of the target detection system will greatly affect the reliability, stability, and security of the entire platform. Therefore, it is urgent to implement a deep learning target detection system stress testing method.

[0004] From a testing perspective, finding real images with different numbers of targets and switching the image types according to the type of task requires a lot of manpower and material resources. In addition, images composed of different targets and backgrounds have different delays for the target recognition algorithm, making it difficult to control variables. The use of artificially synthesized images will be able to better overcome the above shortcomings and achieve an effective evaluation of the stress resistance of the deep learning target detection system.

[0005] Currently, artificially synthesized images are mainly divided into full generation and modification routes. The full generation method mainly relies on artificial intelligence generation algorithms to randomly generate images. The disadvantage is that the generated data varies greatly and it is easy to generate data outside the natural data distribution, which causes inconsistent input delays. The modification-based method mainly modifies the original image through gradient backpropagation (such as adversarial attacks). This method only slightly modifies the original data and can change the judgment results of the artificial intelligence model when the background is almost the same. It can quickly construct diverse extreme scenes on the same photo. Summary of the invention

[0006] The purpose of the present invention is to solve the problem of evaluating the timeliness of target detection systems under extreme conditions. This method evaluates the performance of the algorithm by generating test videos containing different numbers of detection objects and recording the target detection duration. Key technologies include using adversarial attacks to generate multi-target test samples, increasing the number of candidate boxes through iterative optimization, and test process design in practical applications. This method is suitable for evaluating the extreme performance of target detection systems such as drones, industrial control, and scene monitoring, and helps to determine whether the hardware configuration meets the usage requirements or reduce costs.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0008] The present invention provides a deep learning target detection system stress testing method, comprising the following steps:

[0009] Step 1: Input the current image (the first input is the original image) into the tested YOLO model for recognition processing to obtain an image containing potential candidate boxes and filtered candidate boxes;

[0010] Step 1.1: Input the current image into the tested YOLO model, obtain the position of the potential candidate box and the candidate box as the first i Probability of class;

[0011] Step 1.2, filtering the image of the potential candidate frame to obtain the image of the filtered candidate frame;

[0012] Among them, the filter The integrated and The filter function is is the filter function, is the filter function;

[0013]

[0014] Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes , Potential candidate boxes c The probability that the area is the target;

[0015] Use the unconditional threshold T for filtering to obtain the final selected candidate box:

[0016]

[0017] in, For this area i The probability of the target class, i is an integer, is the total number of categories;

[0018] Step 2: Determine the number of candidate boxes after filtering. If the number of candidate boxes is equal to N, stop optimization, add the current image to the test case, and jump to step 4. If it is not equal to N, continue optimization and jump to step 3.

[0019] Step 3: Generate multi-target test samples for the original image to increase the number of potential candidate boxes passing through the filter;

[0020] Step 3.1: For each potential candidate box c Compute the loss function without specifying the target class:

[0021]

[0022] in, is the unconditional threshold, c is a single potential candidate box, The meaning is i The probability of the class target;

[0023] Step 3.2: The total loss function for increasing the number of candidate boxes passing the filter is:

[0024]

[0025] To pass the filter function The candidate box, A potential candidate box predicted for the perturbed image.

[0026] Step 3.3: For each candidate box c Calculate the loss function of the candidate box size constraint ;

[0027] w and h are the width and length of the candidate box respectively;

[0028] Step 3.4: The loss function obtained in step 3.3 Applied to a potential candidate box predicted by the perturbed image , the loss function for the total area constraint of the candidate box is:

[0029]

[0030] Step 3.5: Integrate the total loss function of step 3.2 with the number of candidate boxes passing the filter and the loss function of the total area constraint of the candidate boxes in step 3.4, calculate the perturbation of the entire image according to the loss function, and then update the current image;

[0031] Step 3.5.1. Use loss function Update the image. is the weight coefficient, for , ;

[0032] Step 3.5.2. Loss function Back-propagate the calculated gradient to obtain the perturbation that needs to be added, update the pixels of the image, and jump to step 2 until an image that meets the specified target number is generated;

[0033] Step 4: Perform extreme performance testing on the test case obtained in step 3 to obtain extreme performance test results.

[0034] Step 4.1, input the test images containing different numbers of targets into the target recognition system;

[0035] Step 4.2, record the time difference between input and output;

[0036] Step 4.3: Repeat steps 4.1 and 4.2 until the target number of delays that do not meet the requirements is found, and the extreme performance test result is obtained.

[0037] The present invention also provides a deep learning target detection system stress testing device, comprising:

[0038] The processing unit inputs the current image into the tested YOLO model for recognition processing to obtain an image containing potential candidate frames and filtered candidate frames;

[0039] Determine the unit and determine the number of candidate boxes after filtering. If the number of candidate boxes is equal to N, stop the optimization, add the current image to the test case, and jump to the test unit. If it is not equal to N, continue the optimization and jump to the optimization unit.

[0040] The optimization unit generates multi-target test samples for the original image to increase the number of potential candidate boxes that pass the filter;

[0041] The test unit performs extreme performance test processing on the test cases obtained in the optimization unit to obtain extreme performance test results.

[0042] In the above device, the processing unit includes:

[0043] The position and probability acquisition unit inputs the current image into the tested YOLO model to obtain the position of the potential candidate box and the probability of the candidate box being i Probability of class;

[0044] A filtering processing unit filters the image of the potential candidate frame to obtain an image of the filtered candidate frame;

[0045] Among them, the filter The integrated and The filter function is is the filter function, is the filter function;

[0046] ;

[0047] Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes:

[0048] , Potential candidate boxes c The probability that the area is the target;

[0049] Use the unconditional threshold T for filtering to obtain the final selected candidate box:

[0050]

[0051] in, is the probability that the area is the i-th type target, is the total number of categories.

[0052] In the above device, the optimization unit comprises:

[0053] Loss function calculation unit, for each potential candidate box c Compute the loss function without specifying the target class:

[0054]

[0055] in, is the unconditional threshold, c is a single potential candidate box, The meaning is i The probability of the class target;

[0056] The total loss function calculation unit and the total loss function of increasing the number of candidate boxes passing the filter are:

[0057]

[0058] To pass the filter function The candidate box, A potential candidate box predicted for the image after adding the perturbation;

[0059] Size constraint loss function calculation module, for each candidate box c Calculate the loss function of the candidate box size constraint

[0060] w and h are the width and length of the candidate box respectively;

[0061] Area constraint loss function application module, the loss function obtained by the size constraint loss function calculation module Applied to each candidate box predicted by the system , the loss function for the total area constraint of the candidate box is:

[0062]

[0063] The integrated update module and the integrated total loss function calculation unit add the total loss function of the number of candidate boxes passing the filter and the loss function of the total area constraint of the candidate boxes in the area constraint loss function application module, calculate the perturbation of the entire image according to the loss function, and then update the current image;

[0064] Using loss function Update the image. is the weight coefficient, for , ;

[0065] Loss Function The calculated gradient is back-propagated to obtain the disturbance that needs to be added, and the pixels of the image are updated, jumping to the judgment unit.

[0066] In the above device, the testing unit comprises:

[0067] A test image input submodule inputs test images containing different numbers of targets into a target recognition system;

[0068] The time difference recording submodule records the time difference between input and output;

[0069] The extreme performance test result acquisition submodule is used to repeat the input and recording process until the target number of delays that do not meet the requirements is found, thereby obtaining the extreme performance test result.

[0070] Because the present invention adopts the above technical means, it has the following beneficial effects:

[0071] The present invention aims to evaluate the timeliness of the target detection system under extreme conditions. This method modifies the original video through adversarial attack technology to generate test videos containing different numbers of detection objects, so as to test the extreme performance of the target detection algorithm. The main technical advantages and innovations are as follows:

[0072] 1. Test case generation module: This paper uses adversarial attack technology to modify the original video and generate test videos containing different numbers of detection objects. This method can quickly build diverse extreme scenes while ensuring background consistency, overcoming the difficulty of manpower and material resources and control variables required to find real images with different numbers of targets in traditional methods.

[0073] 2. Extreme performance evaluation module: By recording the real-time detection duration of the target detection algorithm when processing the test video, the present invention can effectively evaluate the timeliness of the algorithm under extreme conditions. This evaluation is crucial to ensure the reliability and stability of the target detection system in practical applications.

[0074] 3. Optimization of NMS algorithm: This paper takes into account the non-maximum suppression (NMS) algorithm commonly used in target detection. By increasing the number of targets in the image and reducing the size of the candidate box through adversarial attacks, the limit of the target increase algorithm is further improved, and the automatic generation of images with different numbers of targets is realized.

[0075] 4. Wide application: The present invention can be applied to target recognition systems such as drones, industrial control, and scene monitoring. By testing the number of extreme targets, it helps determine whether the hardware configuration meets normal usage requirements. It can also be used to reversely determine the available computing hardware configuration, which helps reduce costs.

[0076] 5. Cost and efficiency advantages: Compared with traditional testing methods, the present invention does not require a large number of real images and complex variable control, saving manpower and material costs while improving testing efficiency and accuracy.

[0077] In summary, the method proposed in the present invention not only innovatively utilizes adversarial attack technology to generate test cases, improving the efficiency and accuracy of the test, but also provides new ideas and tools for evaluating and optimizing target detection systems, which is of great significance for promoting the application and development of artificial intelligence technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a simple flow chart of the present invention. DETAILED DESCRIPTION

[0079] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.

[0080] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0081] The present invention mainly targets the target detection system. By constructing the target function, only a small number of original samples are used to generate test cases with different targets. The system delay when the test cases are input into the target detection system is counted to achieve stress testing of the system. The specific technical issues involved include the following aspects:

[0082] 1. Generation of multi-target test samples;

[0083] 2. Target box constraints;

[0084] 3. Practical application of the test method.

[0085] In order to facilitate those skilled in the art to better understand the inventive concept, the following embodiments are described in detail:

[0086] The test cases of the present invention adopt an iterative generation method. First, multi-objective test sample generation and multi-objective box constraints are introduced to introduce the generation targets of each round of iteration. Finally, the actual application of the test method is introduced in detail to illustrate the entire process of iterative generation and testing.

[0087] Multi-target test sample generation

[0088] For the target detection system, the system delay is caused by too many targets. Generally speaking, the targets randomly selected in the original image will not cause too much system delay, so we need to increase the number of targets generated by the test samples. Here we only discuss the case of increasing the number of targets.

[0089] The typical NMS used by YOLO mainly includes the following two steps:

[0090] (1) Use a fixed threshold to filter targets:

[0091]

[0092] (2) Filter using non-conditional thresholds:

[0093]

[0094] in, is the probability that the area is the target, is the threshold value, For this area i The probability of the target class, is the total number of categories. Generally speaking, the candidate box set C Need to pass through the filter:

[0095]

[0096] in Integrated and In order to increase the number of candidate boxes, we need to increase the number of candidate boxes that pass the filter. Therefore, our goal is to increase the number of candidate boxes that are below the threshold. We use adversarial attacks to increase the number of candidate boxes that pass the filter. The loss function of a single candidate box without a specified target class is:

[0097]

[0098] The loss function for a specified target class of a single candidate box is:

[0099]

[0100] The loss function for all candidate boxes is:

[0101]

[0102] is a candidate box that can pass the filter.

[0103] Target box constraints:

[0104] When the number of targets set is very high, the specified number of targets may not be generated by "multi-target test sample generation" alone. In order to improve the potential of the generation method, the multi-target test sample generation method needs to be further enhanced.

[0105] The number of objects generated by "Multi-object Test Sample Generation" varies greatly in size, and large objects will suppress the detection of small objects in the area. Reducing the size of the candidate object box can effectively reduce the IoU (Intersection over Union). Therefore, the loss function is increased to constrain the size of the candidate object box.

[0106]

[0107] c is the candidate box, w and hare the width and length of the candidate box respectively. The loss function for the total area constraint of the candidate box is:

[0108] .

[0109] Practical application of test methods

[0110] (1) Test case generation module

[0111] Select a certain number of original images and set the target number to be generated (such as target number: 20, 21, ..., 49, 50).

[0112] In each round of iteration, the loss function used is:

[0113] , is the weight coefficient, for , .

[0114] In each round of iteration, the loss function back-propagates the calculated gradient and then updates the pixels of the image until the number of generated targets meets the requirement, and the iteration stops.

[0115] The above steps are performed on the selected original images until all images that meet the target number have been generated.

[0116] (2) Extreme performance test module

[0117] The generated test image is input into the target detection system, and the time difference between its input and output is recorded until the number of targets whose delay does not meet the requirement is found.

[0118] Example 1

[0119] The present invention provides a deep learning target detection system stress testing method, comprising the following steps:

[0120] Step 1: Input the current image (the first input is the original image) into the tested YOLO model for recognition processing to obtain an image containing potential candidate boxes and filtered candidate boxes;

[0121] Step 1.1: Input the current image into the tested YOLO model, obtain the position of the potential candidate box and the position of the candidate box as the first i Probability of class;

[0122] Step 1.2: Filter the image of the potential candidate frame to obtain the image of the filtered candidate frame;

[0123] Among them, the filter The integrated and The filter function is is the filter function, is the filter function;

[0124]

[0125] Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes , Potential candidate boxes c The probability that the area is the target;

[0126] Use the unconditional threshold T for filtering to obtain the final selected candidate box:

[0127]

[0128] in, For this area i The probability of the target class, is the total number of categories;

[0129] Step 2: Determine the number of candidate boxes after filtering. If the number of candidate boxes is equal to N (initialized to 10), stop the optimization, add the current image to the test case, and jump to step 4. If it is not equal to N, continue the optimization and jump to step 3.

[0130] Step 3: Generate multi-target test samples for the original image to increase the number of potential candidate boxes passing through the filter;

[0131] Step 3.1: For each potential candidate box c Compute the loss function without specifying the target class:

[0132]

[0133] in, is the unconditional threshold, c is a single potential candidate box, The meaning of is the probability of the i-th type of target;

[0134] Step 3.2: The total loss function for increasing the number of candidate boxes passing the filter is:

[0135]

[0136] To pass the filter function The candidate box, A potential candidate box predicted for the perturbed image.

[0137] Step 3.3: For each candidate box c Calculate the loss function of the candidate box size constraint ;

[0138] w and h are the width and length of the candidate box respectively;

[0139] Step 3.4: The loss function obtained in step 3.3 Applied to a potential candidate box predicted by the perturbed image , the loss function for the total area constraint of the candidate box is:

[0140] Step 3.5: Integrate the total loss function of step 3.2 with the number of candidate boxes passing the filter and the loss function of the total area constraint of the candidate boxes in step 3.4, calculate the perturbation of the entire image according to the loss function, and then update the current image;

[0141] Step 3.5.1. Use loss function Update the image. is the weight coefficient, for , ;;

[0142] Step 3.5.2. Loss function Back-propagate the calculated gradient to obtain the perturbation that needs to be added, update the pixels of the image, and jump to step 2 until an image that meets the specified target number is generated;

[0143] Step 4: Perform extreme performance testing on the test case obtained in step 3 to obtain extreme performance test results.

[0144] Step 4 includes the following steps:

[0145] Step 4.1: Perform extreme performance testing on the test case obtained in step 3 to obtain extreme performance test results.

[0146] Step 4.1.1, input the test images containing different numbers of targets into the target recognition system;

[0147] Step 4.1.2, record the time difference between input and output;

[0148] Step 4.1.3: Repeat steps 4.1.1 and 4.1.2 until the target number of delays that do not meet the requirements is found, and the extreme performance test results are obtained.

[0149] Example 2

[0150] The present invention also provides a deep learning target detection system stress testing device, comprising:

[0151] Processing unit: inputs the current image into the tested YOLO model for recognition processing to obtain an image containing potential candidate frames and filtered candidate frames;

[0152] Judgment unit: judge the number of candidate boxes after filtering. If the number of candidate boxes is equal to N, stop optimization, add the current image to the test case, and jump to the test unit. If it is not equal to N, continue optimization and jump to the optimization unit.

[0153] Optimization unit: Generate multi-target test samples for the original image to increase the number of potential candidate boxes that pass the filter;

[0154] The test unit performs extreme performance test processing on the test cases obtained in the optimization unit to obtain extreme performance test results.

[0155] In the above device, the processing unit includes:

[0156] Position and probability acquisition unit: input the current image into the tested YOLO model, obtain the position of the potential candidate box and the probability of the candidate box being i Probability of class;

[0157] Filtering processing unit: filtering the image of the potential candidate frame to obtain the image of the filtered candidate frame;

[0158] Among them, the filter The integrated and The filter function is is the filter function, is the filter function;

[0159]

[0160] Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes , Potential candidate boxes c The probability that the area is the target;

[0161] Use the unconditional threshold T for filtering to obtain the final selected candidate box:

[0162] in, For this area i The probability of the target class, is the total number of categories.

[0163] In the above device, the optimization unit comprises:

[0164] Loss function calculation unit, for each potential candidate box c Compute the loss function without specifying the target class:

[0165] in, is the unconditional threshold, c is a single potential candidate box, The meaning of is the probability of the i-th type of target;

[0166] The total loss function calculation unit and the total loss function of increasing the number of candidate boxes passing the filter are:

[0167]

[0168] To pass the filter function The candidate box, A potential candidate box predicted for the image after adding the perturbation;

[0169] Size constraint loss function calculation module, for each candidate box c Calculate the loss function of the candidate box size constraint ;

[0170] w and h are the width and length of the candidate box respectively;

[0171] Area constraint loss function application module, the loss function obtained by the size constraint loss function calculation module Applied to add each candidate box , the loss function for the total area constraint of the candidate box is:

[0172] The integrated update module and the integrated total loss function calculation unit add the total loss function of the number of candidate boxes passing the filter and the loss function of the total area constraint of the candidate boxes in the area constraint loss function application module, calculate the perturbation of the entire image according to the loss function, and then update the current image;

[0173] Using loss function Update the image. is the weight coefficient, for , ;

[0174] Loss Function The calculated gradient is back-propagated to obtain the disturbance that needs to be added, and the pixels of the image are updated, jumping to the judgment unit.

[0175] In the above device, the testing unit comprises:

[0176] A test image input submodule inputs test images containing different numbers of targets into a target recognition system;

[0177] The time difference recording submodule records the time difference between input and output;

[0178] The extreme performance test result acquisition submodule is used to repeat the input and recording process until the target number of delays that do not meet the requirements is found, thereby obtaining the extreme performance test result.

Claims

1. A deep learning target detection system stress testing method, characterized by: The following steps are involved: Step 1: Input the current image into the tested YOLO model for recognition processing to obtain an image containing potential candidate boxes and filtered candidate boxes; Step 2: Determine the number of candidate boxes after filtering. If the number of candidate boxes is equal to N, stop optimization, add the current image to the test case, and jump to step 4. If it is not equal to N, continue optimization and jump to step 3. Step 3: Generate multi-target test samples for the original image to increase the number of potential candidate boxes passing through the filter; Step 4: Perform extreme performance test processing on the test case obtained in step 3 to obtain extreme performance test results; Step 1 includes the following steps: Step 1.1: Input the current image into the tested YOLO model, obtain the position of the potential candidate box and the position of the candidate box as the first i Probability of class; Step 1.2: Filter the image of the potential candidate frame to obtain the image of the filtered candidate frame: Among them, the filter For integrated and The filter function is is the filter function, is the filter function; Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes: , Potential candidate boxes c The probability that the area is the target; Use the unconditional threshold T for filtering to obtain the final selected candidate box: ; in, For this area i The probability of the target class, i is an integer, is the total number of categories; Step 3 includes the following steps: Step 3.1: For each potential candidate box c Compute the loss function without specifying the target class: ; in, is the unconditional threshold, c is a single potential candidate box, The meaning is i The probability of class target; Step 3.2: The total loss function for increasing the number of candidate boxes passing the filter is: To pass the filter function The candidate box, A potential candidate box predicted for the image after adding the perturbation; Step 3.3: For each candidate box c Calculate the loss function of the candidate box size constraint , w and h are the width and length of the candidate box respectively; Step 3.4: Substitute the loss function obtained in step 3.3 Applied to a potential candidate box predicted by the perturbed image , the loss function for the total area constraint of the candidate box is: Step 3.5: Integrate the total loss function of step 3.2 to increase the number of candidate boxes that pass the filter and the loss function of step 3.4 to constrain the total area of ​​the candidate boxes, calculate the perturbation of the entire image according to the loss function, and then update the current image; Step 3.5.1: Use loss function Update the image. is the weight coefficient, for , ; Step 3.5.2: Loss Function Back-propagate the calculated gradient to obtain the perturbation that needs to be added, update the pixels of the image, and jump to step 2 until an image that meets the specified target number is generated.

2. A deep learning target detection system stress testing method according to claim 1, characterized in that: Step 4 includes the following steps: Step 4.1: Input the test images containing different numbers of objects into the object recognition system; Step 4.2: Record the time difference between input and output; Step 4.3: Repeat steps 4.1 and 4.2 until the target number of delays that do not meet the requirements is found, and the extreme performance test result is obtained.

3. A deep learning target detection system stress testing device, characterized in that: include: Processing unit: inputs the current image into the tested YOLO model for recognition processing to obtain an image containing potential candidate frames and filtered candidate frames; Judgment unit: judge the number of candidate boxes after filtering. If the number of candidate boxes is equal to N, stop optimization, add the current image to the test case, and jump to the test unit. If it is not equal to N, continue optimization and jump to the optimization unit. Optimization unit: Generate multi-target test samples for the original image to increase the number of potential candidate boxes that pass the filter; Testing unit: Perform extreme performance testing on the test cases obtained in the optimization unit to obtain extreme performance test results; The processing unit includes: Position and probability acquisition unit: input the current image into the tested YOLO model, obtain the position of the potential candidate box and the probability of the candidate box being the first i Probability of class; Filtering processing unit: filtering the image of the potential candidate frame to obtain the image of the filtered candidate frame; Among them, the filter For integrated and The filter function is is the filter function, is the filter function; Use a fixed threshold on the original image for the filter function Filter to get the preliminary selected candidate boxes: , Potential candidate boxes c The probability that the area is the target; Use the unconditional threshold T for filtering to obtain the final selected candidate box: ; in, For this area i The probability of the target class, i is an integer, is the total number of categories; The optimization unit includes: Loss function calculation unit: for each potential candidate box c Compute the loss function without specifying the target class: ; in, is the unconditional threshold, c is a single potential candidate box, The meaning is i The probability of class target; Total loss function calculation unit: The total loss function with increasing number of candidate boxes passing the filter is: To pass the filter function The candidate box, A potential candidate box predicted for the image after adding the perturbation; Size constraint loss function calculation module: for each candidate box c Calculate the loss function of the candidate box size constraint, w and h are the width and length of the candidate box respectively; Area constraint loss function application module: The loss function obtained by the size constraint loss function calculation module Applied to a potential candidate box predicted by the perturbed image , the loss function for the total area constraint of the candidate box is: Integrated update module: The integrated total loss function calculation unit adds the total loss function of the number of candidate boxes passing the filter and the loss function of the total area constraint of the candidate boxes in the area constraint loss function application module, calculates the perturbation of the entire image according to the loss function, and then updates the current image; Using loss function Update the image. is the weight coefficient, for , ; Loss Function The calculated gradient is back-propagated to obtain the disturbance that needs to be added, and the pixels of the image are updated, jumping to the judgment unit.

4. A deep learning target detection system stress testing device according to claim 3, characterized in that: The test unit includes: Test image input submodule: input test images containing different numbers of targets into the target recognition system; Time difference recording submodule: records the time difference between input and output; Extreme performance test result acquisition submodule: used to repeat the input and recording process until the target number of delays that do not meet the requirements is found, and the extreme performance test results are obtained.