Image generation method for neural network model test

Abstract

The invention relates to the field of model testing, and discloses an image generation method for a neural network model test, which comprises the following steps: acquiring a prediction result and confidence distribution of a target neural network on a standard test image set, identifying weaknesses of a model in texture, structure and illumination in combination with a sensitivity analysis algorithm, and obtaining a prediction result of the target neural network; constructing a model error-prone feature set; based on the error-prone feature set, performing semantic-preserving directional disturbance on the original image through a feature parameter injection method; evaluating the structural consistency between the generated image and the original image by using a perception similarity measurement and feature projection algorithm; a self-adaptive strategy is adopted to optimize a generation rule, a model feedback loop mechanism is introduced, and a disturbance parameter range is dynamically adjusted according to the identification error condition of the model on a newly generated image; and automatically generating a model performance diagnosis analysis report in combination with the model output change curve and the weakness dimension labeling result. The method has the advantage of improving the accuracy of model testing.

Description

Technical Field

[0001] The present invention relates to the field of model testing, and particularly to an image generation method for neural network model testing. Background Art

[0002] With the rapid development of deep learning technology, neural network models have been widely used in computer vision tasks such as image classification, object detection, and image segmentation. During the model training and testing process, it often relies on a large amount of well-annotated image data. However, in the model testing stage, in order to evaluate the robustness and generalization ability of the model for different input situations, it is usually necessary to construct a specific test image set. In the prior art, methods such as manual selection, image enhancement, or GAN generation are often used to construct the test set. However, in practical applications, there is a specific problem with existing image generation methods: it is difficult to control the targeted changes of the generated images for the known weaknesses of the model. For example, when the model is sensitive to the edge blurring or brightness offset of a certain type of image, existing image generation tools often cannot accurately control the generation of such features, resulting in insufficient representativeness of the generated images and the test results not being able to truly reflect the performance of the model in edge scenarios. Therefore, it is necessary to design an image generation method for neural network model testing that can improve the accuracy of model testing. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides an image generation method for neural network model testing, which has the advantage of improving the accuracy of model testing and solves the problems in the above background art.

[0004] To achieve the above object of improving the accuracy of model testing, the present invention provides the following technical solutions: An image generation method for neural network model testing, comprising the following steps: Obtain the prediction results and confidence distribution of the target neural network on the standard test image set, combine the sensitivity analysis algorithm to identify the weaknesses shown by the model in texture, structure, and illumination, and construct a model error-prone feature set; Based on the error-prone feature set, perform semantic-preserving directional perturbation on the original image through the feature parameter injection method, and combine fuzzy control to generate a candidate test image set with target feature offsets; Use the perceptual similarity metric and feature projection algorithm to evaluate the structural consistency between the generated image and the original image, and combine the model output difference metric to screen high-response images as effective test samples; Adopt an adaptive strategy to optimize the generation rule, introduce a model feedback loop mechanism, and dynamically adjust the perturbation parameter range according to the recognition error situation of the model on the newly generated images; Combine the model output change curve and the weakness dimension annotation result to automatically generate a model performance diagnosis and analysis report.

[0005] Preferably, the process of constructing the error-prone feature set of the model is as follows: Perform input gradient analysis on the prediction results of the neural network model on the standard test image set, calculate the gradient sensitivity distribution of each input image in different regions, and extract the model's attention area through visualization technology; Divide the image features into three main dimensions: texture, structure, and illumination; Apply controlled perturbations to the test images, record the changes in the model prediction results, quantify the output fluctuations of the model under different feature perturbations; evaluate the response consistency of the model in the local area before and after the perturbation, and identify the unstable response areas; perform cluster analysis on the collected perturbation sensitivity indicators; extract the corresponding texture types, structural patterns, and illumination conditions; identify the typical weak areas and feature parameters, and construct the error-prone feature set of the model.

[0006] Preferably, the process of generating a candidate test image set with target feature offset by combining fuzzy control is as follows: Construct perturbation control parameters according to the error-prone feature set of the model; Construct a fuzzy control system with the perturbation effect feedback as the input and the perturbation parameter adjustment amount as the output; For each perturbation process, calculate the adjustment amplitude of the perturbation parameter through the fuzzy inference mechanism according to the input indicators such as the decrease in the model prediction confidence, structural similarity, and feature offset degree of the image after the current perturbation, and update the perturbation control parameters; Under the guidance of the updated parameters, perform perturbation operations on the original image in a specific direction, and introduce error-prone features without destroying the main semantic structure, so as to generate image samples with target feature offset.

[0007] Preferably, the process of evaluating the structural consistency between the generated image and the original image using the perceptual similarity metric and feature projection algorithm is as follows: Input the original image and the perturbed image into the same pre-trained neural network, and extract the multi-scale semantic features of the intermediate layer; Use the perceptual similarity metric method to compare the feature differences between the original image and the perturbed image; Project the feature vectors of the original image and the perturbed image into the same semantic feature space, analyze their relative positions in the space, and judge the structural stability by calculating the similarity after projection using the Euclidean distance.

[0008] Preferably, the process of judging whether the structure is consistent is as follows: If the Euclidean distance between the original image and the perturbed image in the semantic space is less than or equal to the set distance threshold, it is determined that the structure is consistent; If the Euclidean distance between the original image and the perturbed image in the semantic space is greater than the set distance threshold, it is determined that the structure is inconsistent.

[0009] Preferably, the process of dynamically adjusting the perturbation parameter range according to the recognition error of the model on the newly generated image is as follows: Generate a preliminary set of perturbed images according to the error-prone feature set identified in the early stage and the set perturbation range; Input the preliminarily generated set of perturbed images into the target neural network, obtain the prediction result and the corresponding confidence of each image, and obtain the model output result and confidence; Compare the prediction result of the model with the actual label to calculate the prediction error.

[0010] Preferably, the process of automatically generating a model performance diagnosis and analysis report is as follows: According to the model output data, draw a curve of the model output change to show the output change trend on different perturbed images; Based on the error-prone feature set and feature parameter perturbation, mark the potential weak dimensions of the model; Utilize the curve of the model output change and the annotation result of the weak dimension to automatically generate a diagnosis and analysis report.

[0011] Preferably, the process of screening high-response images is as follows: Calculate the difference degree between the generated image and the original image in terms of the model output result; Set a response threshold, and mark the images whose output change exceeds this threshold as high-response images; Combined with the structural consistency determination result, only retain the image samples with consistent structure but significant response as the final effective test samples.

[0012] Compared with the prior art, the present invention provides an image generation method for neural network model testing, having the following beneficial effects: The image generation method for neural network model testing provided by the present invention can accurately identify the error-prone areas of the model in feature dimensions such as texture, structure, and illumination, generate highly targeted test samples through semantic-preserving directional perturbation and fuzzy control, ensure the structural consistency of the perturbed images by combining perceptual similarity and feature projection algorithms, and then screen high-response images through output differences, significantly improving the effectiveness and diversity of test samples. Introduce a model feedback and adaptive adjustment mechanism, dynamically optimize the perturbation strategy, realize the automatic mining and performance analysis of model weaknesses, and enhance the systematicness, intelligence, and practical value of model testing. Brief Description of the Drawings

[0013] Figure 1 It is a schematic diagram of the method of the present invention. Detailed Embodiments

[0014] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0015] Embodiment 1: Please refer to Figure 1 As shown in the figure, an image generation method for neural network model testing according to an embodiment of the present invention includes the following steps: S1: Obtain the prediction results and confidence distributions of the target neural network on the standard test image set, and combine the sensitivity analysis algorithm to identify the weaknesses shown by the model in terms of texture, structure, and illumination, and construct a model error-prone feature set.

[0016] The process of combining the sensitivity analysis algorithm in S1 to identify the weaknesses shown by the model in terms of texture, structure, and illumination and constructing a model error-prone feature set is as follows: Perform input gradient analysis on the prediction results of the neural network model on the standard test image set, calculate the gradient sensitivity distribution of each input image in different regions, and extract the model attention regions through visualization technology; Divide the image features into three main dimensions: texture, structure, and illumination: Texture: Analyze the responses in high-frequency regions, such as edge-dense, repetitive pattern, and noise-sensitive regions; Structure: Analyze the response fluctuations of the model to the geometric morphology of the image; Illumination: Detect the stability of the model to the illumination conditions such as brightness changes, shadows, and reflections in sensitive regions; Apply controlled perturbations to the test images, record the changes in the model prediction results, and quantify the output fluctuations of the model under different feature perturbations; evaluate the response consistency of the model in the local region before and after the perturbation, use indicators such as structural similarity, perceptual hashing, and local feature point changes to compare the processing differences between the perturbed image and the original image in the same region, and identify the response-unstable regions; perform clustering analysis on the collected perturbation sensitivity indicators, divide the image regions or samples with common sensitivity in specific feature dimensions; extract the corresponding texture types, structural patterns, and illumination condition and other feature parameters; identify the typical weak regions and their feature parameters, and construct a model error-prone feature set, and the feature set describes the sensitivity of the model to specific combinations of image features.

[0017] By obtaining the prediction results and their confidence distributions of the target neural network on the standard test image set, and combining with the sensitivity analysis algorithm, local perturbations and response evaluations are carried out on key perception dimensions such as texture, structure, and illumination in the image to identify the performance weaknesses of the model on specific visual features, thereby constructing a feature set containing typical error-prone patterns. It can accurately mine the instability of the model to complex visual changes, effectively expose potential recognition risks, and provide a basis for the directional generation of subsequent test images and the optimization of model performance.

[0018] S2: Based on the error-prone feature set, perform semantic-preserving directional perturbations on the original image through the feature parameter injection method, and combine fuzzy control to generate a candidate test image set with target feature offsets.

[0019] The process of combining fuzzy control to generate a candidate test image set with target feature offsets in S2 is as follows: According to the model's error-prone feature set, construct perturbation control parameters, such as the noise intensity of texture perturbation, the rotation angle of structure perturbation, and the brightness change amplitude of illumination perturbation, etc. Define an initial perturbation parameter set for each type of image and design corresponding perturbation methods to ensure that the perturbation target is clear and the direction is controllable; Construct a fuzzy control system with the perturbation effect feedback as the input and the perturbation parameter adjustment amount as the output, and formulate fuzzy rules based on expert experience or training data. For example, if the model output change is small and the image similarity is high after perturbation, then appropriately increase the perturbation intensity to achieve dynamic adjustment of the perturbation process; For each perturbation process, according to the input indicators such as the decrease in the model prediction confidence, the structural similarity, and the feature offset degree of the image after the current perturbation, calculate the adjustment amplitude of the perturbation parameters through the fuzzy inference mechanism and update the perturbation control parameters to achieve iterative optimization of the perturbation; Under the guidance of the updated parameters, perform perturbation operations on the original image in a specific direction, and introduce error-prone features without destroying the main semantic structure, thereby generating image samples with target feature offsets.

[0020] Through feature parameter injection based on the model's error-prone feature set, perform semantic-preserving directional perturbations on the original image, and introduce error-prone features in dimensions such as texture, structure, or illumination in a targeted manner without destroying the overall semantic structure of the image; combine the fuzzy control mechanism to dynamically adjust the perturbation intensity and direction, realize the refined adjustment of the perturbation parameters, and then generate a diverse and controllable candidate test image set. While maintaining perceptual consistency, this image set can effectively stimulate the prediction errors of the model on sensitive features, thereby enhancing the coverage and diagnostic ability of the test samples, and improving the accuracy and efficiency of the robustness test and performance analysis of the neural network model.

[0021] S3: Use the perceptual similarity metric and feature projection algorithm to evaluate the structural consistency between the generated image and the original image, and combine the model output difference metric to screen high-response images as valid test samples.

[0022] The process of using the perceptual similarity metric and feature projection algorithm to evaluate the structural consistency between the generated image and the original image in S3 is as follows: Input the original image and the perturbed image into the same pre-trained neural network to extract multi-scale semantic features of the intermediate layer. These features represent the structural information of the image in terms of texture, shape, and semantics, and can be used to measure the consistency between the two images at the perceptual level. Use the perceptual similarity metric method to compare the feature differences between the original image and the perturbed image. The metric can capture the structural changes sensitive to human vision and evaluate whether the perturbation affects the model while maintaining the semantics of the image. Project the feature vectors of the original image and the perturbed image into the same semantic feature space, analyze their relative positions in the space, calculate the similarity after projection, and use the Euclidean distance to judge the structural stability. Let the original image be x and the perturbed image be , input the two into the pre-trained neural network to extract the high-level feature representation vectors. The Euclidean distance formula for the two is: ; In the formula, is the semantic feature vector obtained by extracting the original image through the neural network; is the semantic feature vector obtained by extracting the perturbed image through the neural network; d is the dimension of the feature vector; is the Euclidean distance between the original image and the perturbed image in the semantic space; The process of judging whether the structure is consistent is as follows: If the Euclidean distance between the original image and the perturbed image in the semantic space is less than or equal to the set distance threshold, it is determined that the structure is consistent; If the Euclidean distance between the original image and the perturbed image in the semantic space is greater than the set distance threshold, it is determined that the structure is inconsistent.

[0023] The technical solution of this embodiment is as follows: The perceptual similarity metric and feature projection algorithm are used to comprehensively evaluate the differences between the generated image and the original image at the structural and semantic levels. In the specific process, the perceptual similarity metric is first used to measure the fidelity of the image in terms of structural information, and then the high-dimensional image features are mapped to a low-dimensional space through the feature projection algorithm, so as to visualize and analyze their consistency in the feature space. At the same time, the model output difference is introduced as a response index, and combined with the above evaluation results, the high-response images that have a significant impact on the model are selected as effective test samples. Through the joint strategy of structural consistency evaluation and model response difference measurement, representative test samples can be effectively screened out, highlighting the sensitivity of the model to key feature changes, thereby improving the coverage and discriminative power of the test dataset. This not only improves the robustness of image generation and verification, but also provides accurate and efficient technical support for the interpretability analysis of model performance.

[0024] Embodiment 2: As Figure 1 shown, an image generation method for neural network model testing further includes the following steps: S4: Optimize the generation rule by using an adaptive strategy, introduce a model feedback loop mechanism, and dynamically adjust the perturbation parameter range according to the recognition error situation of the model on the newly generated image.

[0025] The purpose of introducing the model feedback loop mechanism is to dynamically optimize the perturbation generation process, feedback information according to the performance of the model on the newly generated image, and adjust the parameter range of the perturbation in real time, so that the generated perturbed images can more effectively test the weaknesses of the model. By adaptively adjusting the perturbation rule, the quality of the test image set can be improved, ensuring that it can cover the potential weak areas of the model.

[0026] The process of dynamically adjusting the perturbation parameter range according to the recognition error situation of the model on the newly generated image in S4 is as follows: Generate a preliminary perturbed image set according to the error-prone feature set identified in the early stage and the set perturbation range; Input the preliminarily generated perturbed image set into the target neural network, obtain the prediction result and the corresponding confidence of each image, and get the model output result and confidence; Compare the prediction result of the model with the actual label, calculate the prediction error, and the error metric can be the classification error rate or the regression error; Use the prediction error of each perturbed image as feedback information to analyze the recognition error situation of the model. If the perturbed image causes a large prediction error in the model, it means that the perturbation has a greater impact on the model and exposes the weaknesses of the model; Adjust the perturbation parameter range according to the error information feedback by the model. If the perturbed image changes the prediction result of the model, it indicates that the perturbation is effective, and the parameter range of this type of perturbation should be increased. On the contrary, if the perturbation fails to affect the model output, it means the perturbation is ineffective. The adjustment method is as follows: for those perturbed images with larger model errors, increase the perturbation parameter range to make the perturbation more significant; for perturbed images with smaller errors or no changes, reduce the perturbation range to reduce unnecessary perturbations.

[0027] After introducing the model feedback loop mechanism, the perturbation generation process can dynamically adjust the perturbation parameter range according to the error performance of the model on each round of perturbed images. This adaptive optimization strategy ensures that the generated test image set can effectively cover the error-prone areas of the model, while avoiding unnecessary perturbations, thus enhancing the robustness test ability of the model. Through this method, the weaknesses of the model in specific feature dimensions can be more accurately revealed, further enhancing the effectiveness and pertinence of model performance diagnosis.

[0028] S5: Automatically generate a model performance diagnosis analysis report by combining the model output change curve and the weak point dimension annotation results.

[0029] The process of automatically generating a model performance diagnosis analysis report in the above S5 is as follows: Based on the model output data, plot the model output change curve to show the output change trend on different perturbed images. Specifically, the influence of each perturbation feature on the model output at different intensities can be compared. For example, plot the category change curve, confidence change curve or error change curve. The horizontal axis of the curve represents the perturbation intensity, and the vertical axis represents the output change amount. The curve can reveal which features have an impact on the model output and the relationship between the perturbation intensity and the model performance. Based on the error-prone feature set and feature parameter perturbations, label the potential weak point dimensions of the model. The weak point dimensions can be features related to image texture, structure, lighting, etc. By analyzing the model's performance under these features, determine which features the model is vulnerable to perturbations. For example, the model may have a large output difference under the perturbation of lighting changes and a small response to texture perturbations. By annotating the relationship between each perturbation type and the model performance, indicate which features are weak point dimensions. For example, if the model shows instability under lighting changes, it can be labeled as a lighting weakness and relevant analysis can be given. Use the model output change curve and the annotation results of the weak point dimensions to automatically generate a diagnostic analysis report. The report content includes the following aspects: The performance of the model under different perturbations: Analyze the output changes of the model on various perturbed images and plot the change curves.

[0030] Key Weakness Feature Analysis: List the performance weaknesses of the model in different feature dimensions, indicating which feature perturbations have a greater impact on the model.

[0031] Performance Evaluation: Based on the trend of changes in the model output, give the robustness score of the model under different perturbation intensities, as well as possible optimization directions.

[0032] Suggestions and Improvement Measures: According to the diagnostic results, put forward targeted optimization suggestions, such as increasing the training samples of certain features and enhancing the robustness of the model to specific perturbations.

[0033] By combining the model output change curve and the weakness dimension annotation results, the automatically generated performance diagnostic analysis report can clearly and detailedly summarize the robustness performance of the model, help developers quickly identify the potential weaknesses of the model, and provide data support for subsequent optimization. The automated generation process of the report improves the efficiency and accuracy of model performance evaluation, and provides an important basis for the continuous optimization and improvement of the model.

[0034] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0035] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An image generation method for testing a neural network model, characterized in that, It includes the following steps: Obtain the prediction results and confidence distributions of the target neural network on the standard test image set, combine with the sensitivity analysis algorithm to identify the weaknesses shown by the model in terms of texture, structure, and lighting, and construct a model error-prone feature set; Based on the error-prone feature set, perform semantic-preserving directional perturbation on the original image through the feature parameter injection method, and combine with fuzzy control to generate a candidate test image set with target feature offsets; Use the perceptual similarity metric and feature projection algorithm to evaluate the structural consistency between the generated image and the original image, and combine with the model output difference metric to screen high-response images as effective test samples; Adopt an adaptive strategy to optimize the generation rule, introduce a model feedback loop mechanism, and dynamically adjust the perturbation parameter range according to the recognition error of the model on the newly generated images; Combine the model output change curve and the weakness dimension annotation results to automatically generate a model performance diagnostic analysis report.

2. The image generation method for neural network model testing according to claim 1, wherein The process of constructing the model error-prone feature set is as follows: Perform input gradient analysis on the prediction results of the neural network model on the standard test image set, calculate the gradient sensitivity distribution of each input image in different regions, and extract the model attention regions through visualization techniques; Divide the image features into three main dimensions: texture, structure, and lighting; Apply controlled perturbations to the test images, record the changes in the model prediction results, quantify the output fluctuations of the model under different feature perturbations; evaluate the response consistency of the model in the local region before and after the perturbation, and identify the response unstable regions; perform clustering analysis on the collected perturbation sensitivity indicators; extract the corresponding texture types, structural patterns, and lighting conditions; identify the typical weakness regions and feature parameters to construct the model error-prone feature set.

3. An image generation method for neural network model testing according to claim 1, characterized in that, The process of combining fuzzy control to generate a candidate test image set with target feature offsets is as follows: Construct perturbation control parameters according to the model error-prone feature set; Construct a fuzzy control system with the perturbation effect feedback as the input and the perturbation parameter adjustment amount as the output; For each perturbation process, calculate the adjustment amplitude of the perturbation parameters through the fuzzy inference mechanism according to the model prediction confidence decline, structural similarity, and feature offset degree of the image after the current perturbation, and update the perturbation control parameters; Under the guidance of the updated parameters, perform perturbation operations on the original image in a specific direction, and introduce error-prone features without destroying the main semantic structure, so as to generate image samples with target feature offsets.

4. An image generation method for neural network model testing according to claim 1, characterized in that, The process of using the perceptual similarity metric and feature projection algorithm to evaluate the structural consistency between the generated image and the original image is as follows: Input the original image and the perturbed image into the same pre-trained neural network, and extract the multi-scale semantic features of the intermediate layer; Use the perceptual similarity metric method to compare the feature differences between the original image and the perturbed image; Project the feature vectors of the original image and the perturbed image into the same semantic feature space, analyze their relative positions in the space, calculate the similarity after projection, and use the Euclidean distance to judge the structural stability.

5. An image generation method for neural network model testing according to claim 1, characterized in that, The process of judging whether the structure is consistent is as follows: If the Euclidean distance between the original image and the perturbed image in the semantic space is less than or equal to the set distance threshold, it is determined that the structure is consistent; If the Euclidean distance between the original image and the perturbed image in the semantic space is greater than the set distance threshold, it is determined that the structure is inconsistent.

6. The image generation method for neural network model testing according to claim 1, wherein The process of dynamically adjusting the perturbation parameter range according to the recognition error of the model on the newly generated image is as follows: According to the error-prone feature set identified in the early stage and the set perturbation range, a preliminary set of perturbed images is generated; The preliminarily generated set of perturbed images is input into the target neural network to obtain the prediction result and the corresponding confidence level of each image, and the model output result and confidence level are obtained; The prediction error is calculated by comparing the prediction result of the model with the actual label.

7. An image generation method for neural network model testing according to claim 1, characterized in that, The process of automatically generating a model performance diagnosis and analysis report is as follows: According to the model output data, a model output change curve is drawn to show the output change trend on different perturbed images; Based on the error-prone feature set and feature parameter perturbation, the potential weak dimension of the model is marked; Using the model output change curve and the annotation result of the weak dimension, a diagnosis and analysis report is automatically generated.

8. An image generation method for neural network model testing according to claim 1, characterized in that, The screening process of high-response images is as follows: Calculate the difference degree between the generated image and the original image in the model output result; Set a response threshold, and mark the images with output changes exceeding this threshold as high-response images; Combined with the structure consistency determination result, only retain the image samples with consistent structure but significant response as the final effective test samples.

Images