CN117496306A - Multi-level robustness evaluation method and system of machine learning target detection system - Google Patents

Abstract

The invention discloses a multi-level robustness evaluating method and system of a machine learning target detection system, wherein the method comprises the following steps: 1) Defining multi-level critical switching robustness and corresponding critical switching robustness scores for a target detection system; 2) A specific test flow of critical switching robustness; 3) The calculation method of the critical transformation robustness score and the statistics of the test result. Different levels of critical switching robustness may test and measure target detection system robustness under different stringent level requirements. The invention uses image conversion technology to simulate the input image accepted by the target detection system in different real environments, and combines the multi-level critical conversion robustness index evaluation system to cope with the robustness of different real environment condition changes under the condition of carrying out data enhancement on the original test set sample; finally, based on the guidance of the critical switching robustness, the existing target detection system can be repaired and enhanced.

Description

A multi-level robustness evaluation method and system for machine learning target detection systems

Technical field

The invention relates to a multi-level robustness evaluation method and system for a machine learning target detection system, and relates to the technical fields of software engineering and artificial intelligence.

Background technique

In recent years, target detection technology based on machine learning has developed rapidly and has been widely used in various scenarios in society, such as autonomous driving systems. It has brought great convenience to people's lives and work. A target detection model based on machine learning first needs to be trained on a manually labeled training data set to obtain the decision logic for predicting target positions and target categories in images, and then deploy the model to a real environment to receive actual data in the corresponding environment. Input images and perform object detection. However, compared with traditional software systems, the decision-making logic of target detection systems based on machine learning is uninterpretable, so it cannot pass traditional software testing technologies, such as logic coverage, static analysis, etc. Test target detection systems should cope with various deployments Robustness to maintain correct predictions when the environment changes.

When deploying the target detection system to a real environment, the real-time input images received by the system may be affected by different real-world environmental factors and have a large difference in data distribution from the original training set. For example, the difference in light intensity at different times of the day will cause huge differences in the brightness of the image input to the system. When the weather is rainy, the input image will be blurred. The shaking and rotation of the camera will change the angle of the input image. Therefore, the target The system robustness of the detection system in response to different environmental changes has yet to be tested.

Contents of the invention

The purpose of the present invention is to propose a multi-level robustness evaluation method and system for a target detection system to respond to changes in different environments, so as to repair and enhance the existing target detection system based on the evaluation results so that it can be deployed in various in different practical application environments. The present invention uses image conversion technology to simulate the input images accepted by the target detection system in different real-world environments. In the case of data enhancement of the original test set samples, it combines with the multi-level critical conversion robustness index evaluation system to cope with different real-world environmental conditions. Robustness to changes; finally, based on the guidance of critical transition robustness, the existing target detection system can be repaired and enhanced.

In order to achieve the purpose of the present invention, the technical solutions adopted by the present invention are as follows:

A multi-level robustness evaluation method for machine learning target detection systems, the steps of which include:

1) For a target detection system M to be evaluated, input the sample x in the training data set D into the target detection system to obtain the prediction result; the target detection system M is a multi-target detection system;

2) For the prediction result M(x), use transformation T to iteratively transform the sample x. The incremental change of each iterative transformation is δ. At the i+1th transformation, the transformation parameter of transformation T is updated to θ _i+1. = θ _i + δ; the sample x converted in the i-th iteration is T(x; θ _i ), which is input into the target detection system and the prediction result is M(T(x; θ _i )); using multi-level The critical transition robustness measurement index determines whether the prediction result meets the consistency conditions of the prediction results at the corresponding level. When the critical transition robustness judgment at the corresponding level is M(T(x; θ _i ))≠M(x), The conversion parameter θ _i-1 of the conversion T at the i-1th conversion is taken as the critical conversion robustness CTR (x; T, δ of the corresponding level of the target detection system M under the conversion T and the increment δ of the sample x. );

3) Calculate the critical transformation robustness mean value of the corresponding level of the target detection system M under transformation T and increment δ

4) Score based on critical transition robustness Determine the critical conversion robustness score of the corresponding level of the target detection system M under the conversion T and increment δ; where θ _max is the critical conversion robustness CTR (x; T , the maximum value of δ),/> is the average value of the critical transformation robustness CTR (x; T, δ) of the corresponding levels of all samples in the training data set D;

5) Determine the critical conversion robustness of the target detection system M according to the critical conversion robustness scores of each level.

Further, the sample x is an image sample, and the multi-level critical transformation robustness metrics include image-level metrics, class-level metrics and target-level metrics; the corresponding level of critical transformation robustness is CTR(x; T, δ) is the critical transformation robustness at the image level, the critical transformation robustness at the class level, and the critical transformation robustness at the target level.

Furthermore, for the critical transformation robustness at the image level, if the number of target frames in the i-th prediction output changes, it is determined that M(T(x; θ _i ))≠M(x); for the critical transformation at the class level Robustness, if the number of target frames of any category in the i-th prediction output changes, it is determined that M(T(x; θ _i ))≠M(x); for critical conversion robustness at the target level, if If any target frame in the i-th prediction output has a prediction category error or the position deviation is greater than the set value, then it is determined that M(T(x; θ _i ))≠M(x).

Further, to solve the problem of satisfiability of Boolean variables transformed from the output consistency judgment of critical transformation robustness at the target level, the steps include:

1) The prediction result M(x) of the original image in the target detection system M before input conversion is recorded as a triplet set = {(x ₁ , y ₁ , d ₁ ), (x ₂ , y ₂ , d ₂ ), …, (x _n ,y _n ,d _n )}, where n>0, x _i , y _i are the center point coordinates of the i-th target frame, and _di is the maximum allowable center point of the i-th target frame offset distance;

2) The prediction result M (T (x; θ _i )) of the input converted image in the target detection system M is recorded as the point pair set Pairs={(x ₁ , y ₁ ), (x ₂ , y ₂ ),… ,(x _m ,y _m )}, where m>0, x _i , y _i are the center point coordinates of the i-th target frame;

3) The set assigned_pairs of Boolean variables to be solved is initialized to empty;

4) For each element triplet in the pre-conversion prediction result triplet, match an element pair in the post-conversion prediction result Pairs, and the matching condition is Distance(triplet[x,y],pair[x,y]) <=triplet[d], where Distance is the distance calculation function; assign_triplet_{i}_pair_{j}, a Boolean variable representing the result matching

Add to the set assigned_pairs, where i represents the i-th element in Triplets, j represents the j-th element in Pairs, assign_triplet_{i}_pair_{j} represents matching the j-th element in Pairs to the i-th element in Triplets;

5) Use the constraint solver to determine the satisfiability of the Boolean variable set assigned_pairs. If the values of the Boolean variable set can satisfy (SAT), the output consistency at the target level is established; otherwise, if the values of the Boolean variables conflict and cannot be satisfied (UNSAT), the output consistency at the target level is not established.

The specific details of step 4) above are as follows:

4-1) Loop through each element in Triplets;

4-2) For the i-th element triplet in Triplets, traverse every element in Pairs. If the j-th element pair in Pairs can be matched to triplet, then the Boolean variable assign_triplet_{i}_pair_{j} is true and is added to the set assigned_pairs. The Boolean variable assignment implied by this Boolean variable being true includes the i-th element in Triplets. All triples before elements do not match the j-th point pair in Pairs, that is, the list

The values of all Boolean variables in ['assign_triplet_{index}_pair_{j}'foreach index<i] are false, and they are added to the set assigned_pairs. If the j-th element pair in Pairs cannot be matched to triplet, the Boolean variable assign_triplet_{i}_pair_{j} is false and added to the set assigned_pairs;

4-3) Use logical OR operators to combine the Boolean variables constructed by each element in Triplets, and add them to the solving constraints of the constraint solver.

A multi-level robustness evaluation system for a machine learning target detection system, which is characterized by including an image-level critical conversion robustness evaluation module, a class-level critical conversion robustness evaluation module, and a target-level critical conversion robustness evaluation module. Sexual Assessment Module and Comprehensive Assessment Module;

The image-level critical transformation robustness evaluation module is used to iteratively transform the sample x in the training data set D using the image-level transformation T ¹ . The incremental change of each iterative transformation is δ ¹ , the i+1th time During conversion, the conversion parameter of conversion T is updated to θ _i+1 ¹ = θ _i ¹ + δ ¹ ; the sample x converted in the i-th iteration is T ¹ (x; θ _i ¹ ) ¹ , which is input to the target detection system The prediction result M(T ¹ (x; θ _i ¹ )) ¹ is obtained; when the converted output does not meet the requirements of output consistency at the image level, that is, M(T ¹ (x; θ _i ¹ )) ¹ ≠M (x) When ¹ , the conversion parameter θ _i-1 ¹ of conversion T at the i-1th conversion is used as the critical conversion risk of the image level of the target detection system M under conversion T ¹ and increment δ ¹ of sample x. Rodability CTR (x; T ¹ , δ ¹ ), and then calculate the image-level critical transformation robustness mean of the target detection system M under transformation T ¹ and increment δ ¹ Score based on critical transition robustness/> Determine the critical transformation robustness score of the target detection system M at the image level under transformation T ¹ and increment δ ¹ ; where θ _max is the maximum value of the critical transformation robustness of all samples in the training data set D, / > is the average value of the critical transformation robustness of all sample image levels in the training data set D; M(x) ¹ is the prediction result obtained by inputting sample x into the target detection system; the target detection system M is a multi-target detection system;

The class-level critical transformation robustness evaluation module is used to iteratively transform the sample x in the training data set D using the class-level transformation T ² . The incremental change of each iterative transformation is δ ² , the i+1th time During conversion, the conversion parameter of conversion T is updated to θ _i+1 ² =θ _i ² +δ ² ; the sample x converted in the i-th iteration is T ² (x; θ _i ² ) ² , which is input to the target detection system The prediction result M(T ² (x; θ _i ² )) ² is obtained; when the converted output does not meet the requirements of output consistency at the class level, that is, M(T ² (x; θ _i ² )) ² ≠M (x) ² , the conversion parameter θ _i-1 ² of conversion T at the i-1th conversion is used as the critical conversion risk of the class level of the target detection system M under conversion T ² and increment δ ² of sample x. Rodability CTR (x; T ² , δ ² ), and then calculate the class-level critical transformation robustness mean of the target detection system M under transformation T ² and increment δ ² Score based on critical transition robustness/> Determine the critical transformation robustness score of the target detection system M at the class level under transformation T ² and increment δ ² ; where, /> is the average value of critical transformation robustness at the class level of all samples in the training data set D; M(x) ² is the prediction result obtained by inputting sample x to the target detection system;

The target-level critical transformation robustness evaluation module is used to iteratively transform the sample x in the training data set D using the target-level transformation T ^3. The incremental change of each iterative transformation is δ ³ , the i+1th time During conversion, the conversion parameter of conversion T is updated to θ _i+1 ³ =θ _i ³ +δ ³ ; the sample x converted in the i-th iteration is T ³ (x; θ _i ³ ) ³ , which is input to the target detection system The prediction result M(T ³ (x; θ _i ³ )) ³ is obtained; when the converted output does not meet the requirements for output consistency at the target level, that is, M(T ³ (x; θ _i ³ )) ³ ≠M (x) When ³ , the conversion parameter θ _i-1 ³ of conversion T at the i-1th conversion is used as the critical conversion risk of the target level of the target detection system M under conversion T ³ and increment δ ³ of sample x. Robustness CTR (x; T ³ , δ ³ ); then calculate the target level critical conversion robustness mean of the target detection system M under the conversion T ³ and increment δ ³ Score based on critical transition robustness/> Determine the critical conversion robustness score of the target level of the target detection system M under the conversion T ³ and the increment δ ³ ; where, /> is the average value of critical conversion robustness of all sample target levels in the training data set D; M(x) ³ is the prediction result obtained by inputting sample x to the target detection system;

Furthermore, for the critical transformation robustness at the image level, if the number of target frames in the i-th prediction output changes, it is determined that M(T ¹ (x; θ _i ¹ )) ¹ ≠M(x) ¹ ; for the class The critical conversion robustness of the level, if the number of target frames of any category in the i-th prediction output changes, it is determined that M(T ² (x; θ _i ² )) ² ≠M(x) ² ; for the target level ^{The critical} transformation _robustness ^of ) ³ .

Further, to solve the problem of satisfiability of Boolean variables transformed from the output consistency judgment of critical transformation robustness at the target level, the steps include:

1) The predicted result M(x) ³ of the original image in the target detection system M before input conversion is recorded as a triplet set = {(x ₁ , y ₁ , d ₁ ), (x ₂ , y ₂ , d ₂ ) ,...,(x _n ,y _n ,d _n )}, where n>0, x _i , y _i are the center point coordinates of the i-th target frame, and _di is the allowed center point of the i-th target frame Maximum offset distance;

2) Input the predicted result of the converted image in the target detection system M (T ³ (x; θ _i ³ )) ^3, which is recorded as the point pair set Pairs={(x ₁ , y ₁ ), (x ₂ , y ₂ ),…,(x _m ,y _m )}, where m>0, x _i , y _i are the center point coordinates of the i-th target frame;

3) The set assigned_pairs of Boolean variables to be solved is initialized to empty;

4) For each element triplet in the pre-conversion prediction result triplet, match an element pair in the post-conversion prediction result Pairs, and the matching condition is Distance(triplet[x,y],pair[x,y]) <=triplet[d], where Distance is the distance calculation function; assign_triplet_{i}_pair_{j}, a Boolean variable representing the result matching

Add to the set assigned_pairs, where i represents the i-th element in Triplets, j represents the j-th element in Pairs, assign_triplet_{i}_pair_{j} represents matching the j-th element in Pairs to the i-th element in Triplets;

5) Use the constraint solver to determine the satisfiability of the Boolean variable set assigned_pairs. If the value of the Boolean variable set can satisfy (SAT), the output consistency at the target level is established, that is, M(T ³ (x; θ _i ³ )) ³ =

M(x) ³ ; otherwise, the values of the Boolean variables are conflicting and unsatisfiable (UNSAT), and the output consistency at the target level is not established, that is, M(T ³ (x; θ _i ³ )) ³ ≠M(x) ³ .

The specific details of step 4) above are as follows:

4-1) Loop through each element in Triplets;

4-2) For the i-th element triplet in Triplets, traverse every element in Pairs. If the j-th element pair in Pairs can be matched to triplet, then the Boolean variable assign_triplet_{i}_pair_{j} is true and is added to the set assigned_pairs. The Boolean variable assignment implied by this Boolean variable being true includes the i-th element in Triplets. All triples before elements do not match the j-th point pair in Pairs, that is, the list

The values of all Boolean variables in ['assign_triplet_{index}_pair_{j}'foreach index<i] are false, and they are added to the set assigned_pairs. If the j-th element pair in Pairs cannot be matched to triplet, the Boolean variable assign_triplet_{i}_pair_{j} is false and added to the set assigned_pairs;

4-3) Use logical OR operators to combine the Boolean variables constructed by each element in Triplets, and add them to the solving constraints of the constraint solver.

In object detection tasks, the system should make the same prediction results for images with the same semantic information. Based on this criterion, the present invention proposes a multi-level robustness evaluation method for target detection systems, which uses multiple input conversion technologies to evaluate the system's robustness in response to complex environmental changes. The target detection system has multiple independent predicted target frames and categories for an input image, and is a multi-target prediction system. First, for the single-objective prediction system, the critical transformation robustness (CTR, Critical Transformation Robustness) index is defined as follows.

Definition of critical transformation robustness of single-target prediction system: For the neural network M of single-target prediction, the sample x uses transformation T, the iterative incremental transformation parameter θ, θ _i+1 = θ _i + δ, θ ₀ = 0 is the initial value, δ>0, then when θ increases by the change amount δ, assuming that M(T(x; θ _i ))≠M(x) is satisfied, the minimum parameter value is θ _i , and θ _i-1 is the sample The critical transformation robustness CTR(x; T, δ) of x for model M under transformation T and increment δ.

For a data set D, the critical transformation robustness under transformation T and increment δ is defined as the mean critical transformation robustness of all samples in the data set, that is:

The mean represents the overall critical transformation robustness of the samples in the data set D, and is used to measure the overall sensitivity of the data set to input transformations. Based on the definition of critical transition robustness of single target prediction system, the critical transition robustness index of multi-level target detection system is defined as follows.

Definition of multi-level critical transition robustness of target detection system: The target detection system is a multi-target prediction system, which independently predicts different targets in an image at the same time. In view of this characteristic, combined with the critical conversion robustness of the single target prediction system, the critical conversion robustness of the multi-level target detection system is defined as image-level critical conversion robustness, class-level critical conversion robustness and Target level critical transition robustness.

(1) Critical transformation robustness at image level

At this level, the condition for judging that the prediction result of the target detection system has changed (i.e., M(T(x; θ _i ))≠M(x)) is: in the prediction result of the target detection system for an image, the target frame The number has changed. This level of critical conversion robustness satisfies the requirement that the system has the same number of prediction frames for the converted image as for the image before conversion. It is easy to implement and requires less computing resources and execution time. It returns to the single-sample target detection system. The time complexity of processing the results is O(1).

(2) Class-level critical transition robustness

At this level, the condition for judging that the prediction result of the target detection system has changed (i.e., M(T(x; θ _i ))≠M(x)) is: among the prediction results of the target detection system for the image, there is at least one The number of target boxes for categories has changed. The critical transformation robustness of this level meets the requirement that the system has the same number of prediction boxes for the prediction results of the converted image and the pre-conversion image for the prediction results of multiple categories of targets, so it is necessary to traverse the system The prediction results, count the number of prediction boxes for different categories of targets. The implementation difficulty is moderate, the required computing resources and execution time are moderate, and the time complexity of processing the results returned by the single-sample target detection system is O(n).

(3) Target-level critical transition robustness

At this level, the condition for judging that the prediction result of the target detection system has changed (i.e., M(T(x; θ _i ))≠M(x)) is: among the prediction results of the target detection system for the image, there is at least one The prediction of the target box has changed. This change includes a change in the prediction category or a large shift in the position of the prediction box. The critical transformation robustness of this level meets the requirements for the system's prediction results of the converted image and the pre-converted image. For each target frame, in addition to the necessary corresponding transformations in the size and position of the target frame with the conversion, each The predictions before and after a target box transformation should be consistent. Under the requirement of critical conversion robustness, it is necessary to traverse the prediction results of the system when each conversion parameter is incremented, and judge the prediction consistency before and after conversion for each detection frame. It is difficult to implement, requires large computing resources and execution time, and the time complexity of processing the results returned by the single-sample target detection system is O(n ² ). In the specific implementation process of this level of critical conversion robustness, the present invention proposes a method to solve the problem of high difficulty in matching target frames and high computational complexity, which is to model the matching problem of target frames as a satisfiable Boolean variable. Solve problems efficiently and greatly improve calculation speed. The algorithm for solving the satisfiability problem of Boolean variables modeled by the target box matching problem is as follows:

Algorithm 1 Boolean variable satisfiability solution for target frame matching problem modeling

The three levels of critical transformation robustness are sequentially more stringent in judging whether the prediction result of the transformed image is consistent with the prediction result of the pre-transformation image, that is, the critical transformation robustness of the image level is the loosest, and the critical transformation robustness of the class level is more stringent. Robustness is second, and critical transition robustness at the target level is the most stringent. When the sample is converted and the prediction result of the target detection system does not meet the consistency requirements of the prediction results under the critical transformation robustness of the image level, it must not also meet the consistency of the prediction results under the critical transformation robustness of the class level. sexual requirements. And when it does not meet the consistency requirements of prediction results under critical transformation robustness at the class level, it does not necessarily fail to meet the consistency requirements for critical transformation robustness results at the image level. The relationship between critical transition robustness at the target level and critical transition robustness at the class level is similar. The three levels of critical transition robustness can characterize the different intensity robustness requirements of the target detection system, and the corresponding critical transition robustness evaluation method can measure and distinguish the robustness differences of different target detection systems.

After measuring the critical transition robustness of the target detection system under a certain input transformation, in order to further accurately describe the critical transition robustness of the target detection system under a data set D, input transformation T and parameter increment δ, it accounts for the entire As the proportion of the variable range of input transformation parameters on the data set, the present invention proposes the Critical Transformation Robustness Score (CTRS) index, which is defined as follows:

Definition of critical transformation robustness score: The critical transformation robustness score is the ratio of the critical transformation robustness of the input transformation T to the variable range of the parameters of this transformation on the data set D, that is:

The critical transformation robustness score can further intuitively describe the maximum change CTR (D; T, δ) of the sample that the model can withstand under an input transformation and the maximum change that the transformation can bring to the sample. the ratio between. The critical transition robustness score index further makes the evaluation system more complete. The critical transition robustness can be used to evaluate the robustness of different target detection systems under the same environmental change conditions, while the critical transition robustness score can be used It is used to evaluate the robustness of the same system to a variety of different environmental changes.

Multi-level critical transformation robustness During the specific implementation process, since the prediction results of the target detection system will simultaneously predict the detection frame and the category of the target, some input transformation techniques (such as rotation, translation, etc.) will cause target detection The relative position of the frame in the image moves, even causing some target detection frames to move out of the image boundary, which also has varying degrees of impact on the prediction results of the target detection system. Therefore, when judging whether the prediction result of the target detection system changes under a certain level of critical transition robustness under an input transformation that causes the target detection frame to move, the present invention proposes a conservative prediction principle, which is defined as follows.

Conservative prediction principle definition: If the target detection system undergoes input conversion for some predicted target frames in the original image that has not undergone input conversion, and part of the target frame is moved out of the image boundary, then for the converted image, the target frame is allowed Successfully predicted by the target detection system, the target frame is also allowed to be lost by the target detection system. But other unaffected object boxes should be correctly detected and successfully classified.

In order to ensure that the system's robustness to different environmental changes is accurately tested, the test data set D provided to the target detection system needs to meet the consistency and diversity conditions as follows.

Data consistency: For each sample, the number of detection frames and target categories predicted by the target detection system are completely consistent with the label data.

Data diversity: The images in the data set should be as diverse as possible, that is, samples of single targets, multiple targets, multiple categories, and different scenes.

The specific evaluation process of critical transition robustness and critical transition robustness score is shown in Figure 1. First, for the target detection model M to be tested and the test data set D, the input transformation T and its parameter domain [θ _min , θ _max ] and parameter increment δ are determined. The initial critical transition robustness list CTR_List is empty. For all untested samples x∈D, the prediction result of model M for x is M(x), and the conversion parameter θ=θ _min is initialized. The sample x generated after the input transformation T and parameter θ is T(x; θ), and the model’s prediction result for the converted sample is M(T(x; θ)). Iteratively increase the transformation parameter θ with an increment δ until M(x) is inconsistent with M(T(x; θ)) under the critical transformation robustness requirements of a certain level of target detection system, then for sample x after transformation T The critical transition robustness under is denoted as θ-δ. If when the parameter θ has been increased to θ _max , under the critical conversion robustness requirements of a certain level of target detection system, there is still no inconsistency between M(x) and M(T(x; θ)), then for the sample The critical transformation robustness of x under transformation T is denoted by θ _max . Add the critical transition robustness of sample x to the critical transition robustness list CTR_List.

Finally, based on the critical conversion robustness score of the target detection system for a certain input conversion, the evaluation results of the target detection system's robustness ability in responding to certain environmental changes can be obtained.

Compared with the existing technology, the positive effects of the present invention are:

First, a multi-level critical transition robustness index and a critical transition robustness score that can measure the robustness of the target detection system in response to different environmental changes are defined. In response to different robustness requirements, the present invention designs multi-level critical transition robustness indicators and critical transition robustness scores with different computational complexity, which are accurate and effective in measuring the robustness of the target detection system in response to environmental changes. , which solves the problem of robustness evaluation of target detection systems in complex and changeable environments. At the same time, a new method is proposed to solve the problem of high computational complexity of critical transition robustness at the target level, which greatly improves the calculation speed. Finally, combined with the evaluation criteria, the robustness of the target detection system in responding to various real-world environmental changes can be accurately judged, and the upgrade and repair of the target detection system can be guided. At the same time, this method can be well extended to other machine learning model testing frameworks and has good scalability.

Description of drawings

Figure 1 is a test flow chart of the critical transition robustness score of a target detection system under input transition.

Detailed ways

The present invention will be further described in detail below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, those skilled in the art will make modifications to various equivalent forms of the present invention and all fall within the scope defined by the appended claims of this application.

This example takes the mask target detection system based on YoLoV5 as an example to give the evaluation process of multi-level critical transition robustness. This target detection system predicts face target frames and mask target frames that exist in an image.

1. Input conversion selection and parameter settings

This example uses the two input conversion technologies of brightness change and rotation to further explain the test process in detail. The test parameter settings of the input conversion are shown in Table 1. The input conversion technology used by the present invention to simulate various changes in the real environment includes but is not limited to:

(1) Brightness change simulates the situation where the brightness of the image data input to the target detection system changes due to changes in the intensity of the light source in the environment (such as the intensity of sunlight at different times of the day, changes in the power of lighting lamps, etc.).

(2) Image blurring and image quality compression can simulate the degradation of the image data input to the target detection system due to changes in weather in the environment (such as rain, fog, etc.) or input device problems, resulting in blurring of the target object. clear situation.

(3) Changing the contrast can simulate the situation where the contrast of the images input to the target detection system is different due to different configurations of the devices when using different input devices.

(4) Rotation can simulate the situation where the angle of the image input to the target detection system changes significantly due to certain reasons (such as tilt of the input device, etc.).

(5) Translation can simulate the situation where the target object in the image appears in different relative positions in the image.

(6) Horizontal miscutting can simulate the situation where the input device is placed in a different position, resulting in a different viewing angle of the image input into the target detection system.

(7) Image scaling can simulate the situation where the distance between the target and the input device is different, resulting in a different proportion of the target in the entire image.

Table 1 Input conversion and parameter settings

The settings of the maximum and minimum values of the input conversion parameters need to be specifically formulated based on the specific input conversion technology, parameter variable range, and the actual needs of the test party. Under the critical conversion robustness standard of the target level, the present invention uses the input of the coordinates of the center point of the prediction frame of the transformed image and the center point of the corresponding prediction frame of the original image to determine whether the position of the target frame has shifted significantly. The distance between the converted mapping coordinates is greater than half of the diagonal distance of the corresponding target frame in the original image. This judgment standard can be adapted according to different requirements of the actual system.

2. Calculation of critical transition robustness score and statistics of test results

The target detection model M is on the test data set D. The critical transformation robustness with respect to the input transformation T is the mean of the critical transformation robustness of all samples in D. The initial average parameter of the test data set D for the input transformation T is θ _avg , then the critical transformation robustness score of the model M on the test data set D with respect to the input transformation T is recorded as

On the YoLoV5 architecture mask target detection system, the critical conversion robustness test results of the target detection system at three different levels for the two input conversions of brightness change and rotation are shown in Table 2, Table 3 and Table 4.

Table 2 Critical transformation robustness at image level

Table 3 Critical transition robustness at class level

Table 4 Critical transition robustness at target level

Experimental results show that the three levels of critical conversion robustness are sequentially more stringent in judging whether the prediction result of the converted image is consistent with the prediction result of the pre-conversion image, that is, the critical conversion robustness of the image level is the loosest, and the class The robustness of critical transformations at the image level is second, and the critical transformation robustness at the image level is the most stringent. When a sample undergoes input conversion and the prediction result of the target detection system does not meet the prediction result consistency requirements under the critical conversion robustness at the image level, the prediction results under the critical conversion robustness at the class level must not meet the prediction results at this time. Consistency requirements. And when it does not meet the consistency requirements of prediction results under critical transformation robustness at the class level, it does not necessarily fail to meet the consistency requirements for critical transformation robustness results at the image level. As shown in the experimental results, the critical transformation robustness of the class level is equal to or smaller than the critical transformation robustness of the image level. Among them, the critical transformation robustness of the right-rotated class level is significantly smaller than the critical transformation robustness of the image level. This shows that the image level only focuses on whether the number of target frames in the prediction results has changed, without considering whether the prediction categories of the target frames are consistent, while the critical transformation robustness of the class level focuses on whether the number of prediction frames has changed. Also pay attention to whether the target detection system's prediction category for the detection frame has also changed. Therefore, when the number of predicted target boxes in the image has not changed but the predicted categories have changed due to the input transformation technology, the class-level critical transformation robustness can be a good measure of the system's robustness capability at this time. But at the same time, the critical transformation robustness at the class level has higher testing time requirements than the critical transformation robustness at the image level. The relationship between critical transition robustness at the target level and critical transition robustness at the class level is similar.

In summary, the invention proposes a robustness testing method for a machine learning target detection system based on input conversion technology, which can accurately and effectively test the robustness of the target detection system in response to different environmental changes, and is applicable for various machine learning models.

Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smart phone, etc.), which includes a memory and a processor, the memory stores a computer program, and the computer program is configured to be The computer program includes instructions for executing each step of the method of the present invention.

Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM/RAM, magnetic disk, optical disk). The computer-readable storage medium stores a computer program. When the computer program is executed by a computer, , implement each step of the method of the present invention.

Other embodiments of the invention:

The present invention does not limit the input conversion technology adopted and its corresponding parameter range and parameter increment, as well as the data set required for testing.

The present invention does not limit the specific implementation of three different levels of critical transition robustness determination for target detection and the specific classification of critical transition robustness levels.

The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and implement it accordingly. Those of ordinary skill in the art can understand that various substitutions, changes and modifications can be made without departing from the spirit and scope of the present invention. Modifications are possible. The present invention should not be limited to the contents disclosed in the embodiments of this specification. The protection scope of the present invention shall be defined by the claims.

Claims (10)

1. A multi-level robustness evaluating method of a machine learning target detection system comprises the following steps:

1) For a target detection system M to be evaluated, inputting a sample x in a training data set D into the target detection system to obtain a prediction result; the target detection system M is a multi-target detection system;

2) For the prediction result M (x), performing iterative conversion on the sample x by using a conversion T, wherein the increment change of each iterative conversion is delta, and the conversion parameter of the conversion T is updated to theta when the (i+1) th conversion is performed _i+1 ＝θ _i +δ; sample x after the ith iteration conversion is T (x; theta) _i ) Is input into the target detection system to obtain a prediction result of M (T (x; theta) _i ) A) is provided; judging whether the predicted result meets the condition of the predicted result consistency under the corresponding level by using the multi-level critical conversion robustness measure index, and judging M (T (x; theta) _i ) When M (x) is not equal to the i-1 st conversion parameter θ of the conversion T _i-1 Critical transition robustness CTR (x; T, δ) as sample x for the corresponding level of the target detection system M at transition T and increment δ;

3) Calculating a critical conversion robustness mean value of the corresponding level of the target detection system M under the conversion T and the increment delta

4) Robustness score based on critical transitionsDetermining a critical conversion robustness score of a corresponding level of the target detection system M under the conversion T and the increment delta; wherein θ _max Maximum value of critical transition robustness CTR (x; T, delta) for the corresponding level of all samples in training data set D, +.>The average value of the critical conversion robustness CTR (x; T, delta) of the corresponding level of all samples in the training data set D;

5) And determining the critical switching robustness of the target detection system M according to the critical switching robustness scores of the levels.

2. The method of claim 1, wherein the sample x is an image sample, and the multi-level critical transition robustness metrics include an image level metric, a class level metric, and a target level metric; the critical switching robustness CTR (x; T, δ) of the corresponding level is the critical switching robustness of the image level, the critical switching robustness of the class level, and the critical switching robustness of the target level.

3. The method according to claim 2, wherein for the critical transition robustness at the image level, if the number of target frames of the i-th prediction output is changed, it is determined that M (T (x; θ _i ) M (x); for critical transition robustness at class level, if the number of target frames of any class of the ith prediction output is changed, it is determined that M (T (x; θ) _i ) M (x); for the critical transition robustness of the target level, if any target frame of the ith prediction output has a prediction type error or a positional shift greater than a set value, it is determined that M (T (x; θ) _i ))≠M(x)。

4. A method according to claim 1, 2 or 3, wherein the transformation T includes, but is not limited to, an image transformation such as a luminance transformation, a rotation angle transformation, etc. on the sample x.

5. A method according to claim 3, characterized in that for a target level of critical transition robustness, M (T (x; θ _i ) The method for judging whether M (x) is consistent is as follows:

1) The prediction result M (x) of the original image before input conversion in the target detection system M is recorded as a triplet set triplets= { (x) ₁ ,y ₁ ,d ₁ ),(x ₂ ,y ₂ ,d ₂ ),…,(x _n ,y _n ,d _n ) N is }, where n>0，x _i ,y _i Is the center point coordinate of the ith target frame, d _i The maximum offset distance of the center point allowed by the ith target frame;

2) The prediction result M (T (x; θ) of the converted image in the target detection system M is input _i ) Denoted as point pair set Pairs = { (x) ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _m ,y _m ) -wherein m>0，x _i ,y _i The center point coordinates of the ith target frame;

3) Initializing a set of assigned_pairs of Boolean variables to be solved as null;

4) For each element triplet in the triplet triples of pre-conversion prediction results, matching an element pair in a pair of post-conversion prediction results, wherein the matching condition is Distance (triplet [ x, y ], pair [ x, y ]) < = triplet [ d ], and Distance is a Distance calculation function; adding a Boolean variable assignment_triple_i_pair_j representing result matching to the set assignment_pair, wherein i represents an ith element in the triple, j represents a jth element in the pair, and assignment_triple_i_pair_j represents matching the jth element in the pair to the ith element in the triple;

5) Determining the satisfiability of the Boolean variable set assigned_pairs by using a constraint solver; if the value of the Boolean variable set can be satisfied, the consistency of the output at the target level is established, i.e., M (T (x; θ) _i ) M (x); otherwise, if the value of the Boolean variable has conflict, the output consistency under the target level is not satisfied, namely

M(T(x；θ _i ))≠M(x)。

6. The multi-level robustness evaluation system of the machine learning target detection system is characterized by comprising an image-level critical transformation robustness evaluation module, a class-level critical transformation robustness evaluation module, a target-level critical transformation robustness evaluation module and a comprehensive evaluation module;

image level critical transition robustness assessment module for using image level transition T ¹ Performing iterative conversion on samples x in the training data set D, wherein the incremental change of each iterative conversion is delta ¹ The conversion parameter of the conversion T at the i+1th conversion is updated to θ _i+1 ¹ ＝θ _i ¹ +δ ¹ The method comprises the steps of carrying out a first treatment on the surface of the Sample x after the ith iterative conversion is T ¹ (x；θ _i ¹ ) ¹ Input it to the target detection system to obtain the prediction result M (T ¹ (x；θ _i ¹ )) ¹ The method comprises the steps of carrying out a first treatment on the surface of the When the converted output does not meet the requirement of output consistency at the image level, i.e. M (T ¹ (x；θ _i ¹ )) ¹ ≠M(x) ¹ At that time, the conversion parameter theta of the i-1 st conversion time conversion T is calculated _i-1 ¹ As sample x at transition T for the target detection system M ¹ And delta ¹ Critical transition robustness CTR (x; T) at lower image level ¹ ,δ ¹ ) Then calculate the target detection system M at transition T ¹ And delta ¹ Critical transition robustness mean at lower image levelAccording to the critical switching robustness score +.> Determining that the target detection system M is in transition T ¹ And delta ¹ A lower image level critical transition robustness score; wherein θ _max For the maximum of the critical switching robustness of all samples in the training dataset D +.>The average value of the critical conversion robustness of all sample image levels in the training data set D is obtained; m (x) ¹ Inputting a prediction result obtained by the target detection system for a sample x; the target detection system M is a multi-target detection system;

a class-level critical transition robustness assessment module for using class-level transitions T ² Performing iterative conversion on samples x in the training data set D, wherein the incremental change of each iterative conversion is delta ² The conversion parameter of the conversion T at the i+1th conversion is updated to θ _i+1 ² ＝θ _i ² +δ ² The method comprises the steps of carrying out a first treatment on the surface of the Sample x after the ith iterative conversion is T ² (x；θ _i ² ) ² Input it to the target detection system to obtain the prediction result M (T ² (x；θ _i ² )) ² The method comprises the steps of carrying out a first treatment on the surface of the When the converted output does not meet the requirement of output consistency at class level, i.e. M (T ² (x；θ _i ² )) ² ≠M(x) ² At that time, the conversion parameter theta of the i-1 st conversion time conversion T is calculated _i-1 ² As sample x at transition T for the target detection system M ² And delta ² Critical transition robustness CTR (x; T) at the class-down level ² ,δ ² ) Then calculate the target detection system M at transition T ² And delta ² Critical transition robustness at the class-down level Value ofAccording to the critical switching robustness score +.>Determining that the target detection system M is in transition T ² And delta ² Critical transition robustness scores at the class-down level; wherein,

the average value of the critical conversion robustness of all sample class levels in the training data set D is obtained; m (x) ² Inputting a prediction result obtained by the target detection system for a sample x;

a target level critical transition robustness assessment module for using target level transitions T ³ Performing iterative conversion on samples x in the training data set D, wherein the incremental change of each iterative conversion is delta ³ The conversion parameter of the conversion T at the i+1th conversion is updated to θ _i+1 ³ ＝θ _i ³ +δ ³ The method comprises the steps of carrying out a first treatment on the surface of the Sample x after the ith iterative conversion is T ³ (x；θ _i ³ ) ³ Input it to the target detection system to obtain the prediction result M (T ³ (x；θ _i ³ )) ³ The method comprises the steps of carrying out a first treatment on the surface of the When the converted output does not meet the requirement of output consistency at the target level, i.e. M (T ³ (x；θ _i ³ )) ³ ≠M(x) ³ At that time, the conversion parameter theta of the i-1 st conversion time conversion T is calculated _i-1 ³ As sample x at transition T for the target detection system M ³ And delta ³ Critical transition robustness CTR (x; T) at lower target level ³ ,δ ³ ) The method comprises the steps of carrying out a first treatment on the surface of the Then calculate the target detection system M at transition T ³ And delta ³ Critical transition robustness mean of lower target levelAccording to the critical switching robustness score +. > Determining that the target detection system M is in transition T ³ And delta ³ A critical transition robustness score for the lower target level; wherein (1)>The average value of the target level critical conversion robustness of all samples in the training data set D is obtained; m (x) ³ Inputting a prediction result obtained by the target detection system for a sample x;

the comprehensive evaluation module is used for converting the robustness score CTRS (D; T) according to the critical of each level ¹ ,δ ¹ )、CTRS(D；T ² ,δ ² )、CTRS(D；T ³ ,δ ³ ) The critical switching robustness of the object detection system M is determined.

7. The system of claim 6, wherein for critical transition robustness at the image level, if the number of target frames of the ith prediction output changes, determining M (T ¹ (x；θ _i ¹ )) ¹ ≠M(x) ¹ The method comprises the steps of carrying out a first treatment on the surface of the For critical transition robustness at class level, if the number of target frames of any class of the ith prediction output changes, then it is determined that M (T ² (x；θ _i ² )) ² ≠M(x) ² The method comprises the steps of carrying out a first treatment on the surface of the For the critical transition robustness of the target level, if any target frame of the ith prediction output has a prediction type error or a position shift greater than a set value, it is determined that M (T ³ (x；θ _i ³ )) ³ ≠M(x) ³ 。

8. The system of claim 7, wherein for a target level of critical transition robustness, M (T ³ (x；θ _i ³ )) ³ 、M(x) ³ The method for judging whether the two methods are consistent is as follows:

1) Inputting the predicted result M (x) of the original image before conversion in the target detection system M ³ Recorded as triplet sets three= { (x) ₁ ,y ₁ ,d ₁ ),(x ₂ ,y ₂ ,d ₂ ),…,(x _n ,y _n ,d _n ) N is }, where n>0，x _i ,y _i Is the center point coordinate of the ith target frame, d _i The maximum offset distance of the center point allowed by the ith target frame;

2) Inputting the prediction result M (T ³ (x；θ _i ³ )) ³ Denoted as point pair set Pairs = { (x) ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _m ,y _m ) -wherein m>0，x _i ,y _i The center point coordinates of the ith target frame;

3) Initializing a set of assigned_pairs of Boolean variables to be solved as null;

4) For each element triplet in the triplet triples of pre-conversion prediction results, matching an element pair in a pair of post-conversion prediction results, wherein the matching condition is Distance (triplet [ x, y ], pair [ x, y ]) < = triplet [ d ], and Distance is a Distance calculation function; adding a Boolean variable assignment_triple_i_pair_j representing result matching to the set assignment_pair, wherein i represents an ith element in the triple, j represents a jth element in the pair, and assignment_triple_i_pair_j represents matching the jth element in the pair to the ith element in the triple;

5) Determining the satisfiability of the Boolean variable set assigned_pairs by using a constraint solver; if the value of the Boolean variable set can be satisfied, the consistency of the output at the target level is established, namely M (T ³ (x；θ _i ³ )) ³ ＝M(x) ³ The method comprises the steps of carrying out a first treatment on the surface of the Otherwise, if the value of the Boolean variable has conflict, the output consistency at the target level is not satisfied, namely M (T ³ (x；θ _i ³ )) ³ ≠M(x) ³ ；

The specific details of the above step 4) are as follows:

4-1) looping through each element in the Triplets;

4-2) traversing each element in the pair for the ith element triplet in the Triplets; if the jth element pair in the pair can be matched with the triplet, the Boolean variable assignment is true, the Boolean variable assignment contained in the Boolean variable assignment is false, and the Boolean variable assignment comprises that all Triplets before the ith element in the triplet are not matched with the jth point pair in the pair, namely, all Boolean variables in a list [ ' assignment_triplet { index } ' pair } ' for index < i ] are false, and the Boolean variable assignment is added into the set assigned_pair; if the jth element pair in the pair cannot be matched with the triplet, the Boolean variable assignment_triplet_ { i } -pair_ { j } is false, and the value is added into the set assignment_pair;

4-3) combining the Boolean variables constructed for each element in the Triplets using logical or operators, and adding them to the solution constraints of the constraint solver.

9. A server comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for performing the steps of the method of any of claims 1 to 5.

10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 5.