A target detection network evaluation method based on quantification evaluation and interpretability

Abstract

The present application relates to a target detection network evaluation method based on quantitative evaluation and explainability method. The present application combines quantitative evaluation and explainability method, which can more comprehensively and accurately evaluate the performance and decision basis of the target detection network. The present application uses multiple quantitative evaluation indicators (accuracy, recall rate, average precision, FPS) to comprehensively evaluate the target detection network. An explainability method based on perturbation and class activation mapping is proposed, which can help users understand the decision-making process of the target detection network. A visualization method based on gradient descent method and regularization technique is proposed, which can help users understand the semantics contained in the target detection network. The explainability method proposed in the present application can intuitively show the internal operation process and knowledge of the target detection network, thereby enhancing the explainability of the model and improving the credibility of the model results. Experimental results show that the present application can effectively evaluate the performance of the target detection network, improve the explainability of the model, and can be used for the design, evaluation and improvement of the target detection network.

Description

Technical Field

[0001] This invention relates to target detection networks, quantitative evaluation metrics, and neural network interpretability techniques, and particularly to a target detection network evaluation method based on quantitative evaluation and interpretability methods. Background Technology

[0002] With the rapid development of deep learning, artificial intelligence technology has been gradually applied in various fields, and has been widely developed and applied in areas such as license plate detection and autonomous driving. Early object detection methods were mainly based on image processing techniques, such as threshold segmentation and edge detection. These methods have limitations, such as being sensitive to interference from lighting and noise, and being unable to handle complex scenes and multi-object detection problems. Traditional object detection algorithms, due to their sliding window model, have a certain degree of simplification in feature extraction and matching, and their adaptability, accuracy, and detection speed are relatively poor. Deep learning-based object detection algorithms utilize the powerful feature extraction and generalization capabilities of neural networks. They only require a simple network structure to achieve advantages in efficiency and accuracy, thus quickly surpassing traditional object detection algorithms and becoming the most mainstream object detection algorithm.

[0003] Quantitative evaluation of object detection networks is an important means of objectively and systematically assessing their performance. The development of object detection technology requires a unified standard to measure the merits of different algorithms. Through quantitative evaluation, the performance of different algorithms can be compared, performance bottlenecks can be identified, and thus algorithm optimization can be guided. Quantitative evaluation of object detection networks is a crucial means of objectively assessing their performance, reliability, and applicability. It helps researchers find better models and algorithms, improving the effectiveness of object detection algorithms in practical applications.

[0004] However, existing deep neural networks are typically black-box models, and their highly nonlinear nature makes it difficult to understand the internal decision-making processes. Therefore, using only quantitative evaluation metrics cannot fully meet users' requirements for the safety and reliability of object detection algorithms when deployed in real-world task scenarios.

[0005] The inherent lack of interpretability in deep neural networks hinders the practical application of deep neural network-based object detection algorithms. To address this issue, interpretability methods are needed to demystify the black-box model, demonstrating its decision-making process in human-understandable terms. Interpretability methods have diverse applications: they can help users understand the decision-making process and rationale behind object detection networks, correct errors in object detection algorithms, and improve their performance. Currently, interpretability technology has made significant progress, resulting in a series of related technical approaches. Summary of the Invention

[0006] To address, or at least partially address, the aforementioned technical problems, this invention provides a target detection network evaluation method based on quantitative evaluation and interpretability methods.

[0007] This invention proposes a novel interpretability method for deep learning-based object detection networks. Current evaluations of object detection networks primarily focus on quantitative assessments, evaluating model performance through metrics such as accuracy and FPS. While this approach provides quantitative results, it fails to help users understand the model's decision-making process and outcomes, thus limiting its application. To address this issue, this invention introduces interpretability methods into the evaluation process of object detection networks. By explaining and demonstrating the model's decision-making process and results, a deeper understanding of the model's decision-making basis and outcomes can be gained, improving interpretability and better meeting the needs of practical applications. The proposed evaluation method employs multiple interpretability techniques, including a saliency map method based on perturbation and class activation mapping, a semantic visualization method based on gradient descent and regularization, and corresponding interpretation analysis methods. This invention analyzes and explains the model's decision-making process and results, presenting the explanations in the form of saliency maps, synthesized images, etc., enabling others to better understand and use the evaluation results.

[0008] This invention provides an evaluation method for target detection networks based on quantitative evaluation and interpretability methods. The evaluation method includes the following steps:

[0009] Step S1: Determine the evaluation objectives and indicators, and clarify the target detection network and relevant evaluation indicators to be evaluated;

[0010] Step S2: Data acquisition and preprocessing. Collect images and preprocess them to construct a training set for training the model, a test set for evaluating model performance, and samples for generating interpretations.

[0011] Step S3: Train the selected object detection network on the training set;

[0012] Step S4: Evaluate the object detection network using the test set and calculate quantitative evaluation metrics, including accuracy, recall, precision, F1 score, etc.

[0013] Step S5: Propose an interpretability method based on perturbation and class activation mapping to interpret the target detection network, analyze the interpretability of the target detection network, and calculate the interpretability score;

[0014] Step S6: Propose a visualization technique based on nonnegative matrix factorization and gradient descent to explain the target detection network and analyze the semantics of the network structure in the target detection network;

[0015] Step S7: Analyze and present the evaluation results.

[0016] Thus, the evaluation method can not only quantitatively evaluate deep learning-based object detection networks, but also interpret these networks using interpretable methods, enabling more efficient use of object detection networks.

[0017] The beneficial effects of this invention are: it enables the evaluation of object detection networks based on quantitative assessment and interpretability methods, providing a technical foundation for interpretability methods and quantitative assessment of interpretation in deep learning networks. The technical approach provided by this invention can offer empirical methods for evaluating object detection networks. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Appendix Figure 1 A flowchart illustrating a target detection network evaluation method based on quantitative evaluation and interpretability methods provided in this embodiment of the invention;

[0021] Appendix Figure 2 The results of the mAP metric evaluation of the target detection network in this embodiment of the invention are first calculated by using different standards (different IoU thresholds) for the predicted box and the ground box or focusing on different types of objects (different areas), and then the detection accuracy of different categories of objects (such as aeroplane, bicycle, bird, etc.) is calculated.

[0022] Appendix Figure 3 The results of the interpretation method based on perturbation and class activation mapping proposed in this invention are shown in the comparison images. It can be seen that the thermal image regions generated by this algorithm are more concentrated and have less noise, with more accurate localization. The interpretation results of D-RISE and Grad-CAM not only highlight too many irrelevant regions, but also have poor class discrimination and localization effects, making them misleading.

[0023] Appendix Figure 4The interpretability score calculation results based on perturbation and class activation mapping proposed in this embodiment of the invention are shown. During the deletion process, as more and more important pixels are removed, this index measures the decrease in the predicted class probability, and then calculates the degree of change in the predicted probability. Its intuitive meaning is that a good interpretation result should be able to accurately highlight the pixels or regions that play an important role in the prediction, so removing them can significantly change the model's predicted probability;

[0024] Appendix Figure 5 The semantic visualization method proposed in this embodiment of the invention displays the visualization results of the extended weights, with different colors (red for positive and blue for negative) representing the working mode of the extended weights;

[0025] Appendix Figure 6 The semantic visualization method proposed in this invention displays the visualization results of neuronal semantics, and the properties (color, texture, etc.) of the synthesized image can convey the semantics of the neuron to the user. Detailed Implementation

[0026] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0028] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0029] This invention proposes a target detection network evaluation method based on quantitative evaluation and interpretability, focusing on the evaluation system of target detection networks. The evaluation method includes the following steps:

[0030] Step S1: Determine the evaluation objectives and metrics, and clarify the target detection network to be evaluated and the relevant evaluation metrics. Evaluation objectives typically include aspects such as model performance, robustness, generalization ability, and decision-making basis, while evaluation metrics are specific quantitative indicators that can measure some of the evaluation objectives.

[0031] Step S1.1: When determining the evaluation target, it is necessary to select it based on the actual application requirements.

[0032] In industrial production, object detection networks need to be applied to real-world scenarios, requiring consideration of metrics such as model accuracy, recall, and false positive rate. In the field of autonomous driving, robustness and generalization ability of the model are crucial, enabling it to adapt to complex conditions including varying weather, lighting, and scene changes.

[0033] Step S1.2: When determining the evaluation indicators, it is necessary to consider the specific measurement method of the evaluation objectives.

[0034] For model precision and recall metrics, accuracy, precision, and recall can be used as indicators; for false positive rate metrics, the number of false positives and false positive rate can be used as indicators; and for robustness and generalization ability metrics, evaluation methods such as adversarial example testing and dataset transfer can be used, with the model's performance on the new dataset as the metric. Step S2: Data Acquisition and Preprocessing: Collect images and preprocess them to construct a training set for training the model, a test set for evaluating model performance, and samples for generating interpretations. The training set should be large enough to include a sufficient number of task scenarios to improve the model's working ability and generalization ability. Correspondingly, the test set should also contain a sufficient number of representative samples from task scenarios to fully test the model's working ability and generalization ability.

[0035] Step S2.1: Determine the content and size of the dataset.

[0036] When determining the dataset, it is necessary to consider the application scenario and target of the object detection network to collect appropriate data. The dataset size should not be too small when training a deep neural network-based object detection network, and the diversity of image features within the dataset should also be taken into account.

[0037] Step S2.2: Data acquisition and preprocessing.

[0038] Collecting a certain number of images can be done in various ways, such as acquiring them from the internet, taking photos on-site, or using existing datasets. When collecting images, it's important to consider factors such as image quality and the balance of object categories. Typically, a series of preprocessing steps are also required, such as resizing images to a uniform size, cropping images, standardizing image data, and data augmentation.

[0039] Step S2.3: Label the image and save it.

[0040] When labeling images using the correct software, object detection tasks typically use bounding boxes as annotations. The labeled samples need to adhere to certain standards and requirements, such as the coverage of the object by the bounding boxes and the minimum spacing between them. During the annotation process, it is crucial to maintain accuracy and consistency. After annotation is complete, save the images and annotation results in the correct format for further work.

[0041] Step S2.4: Construct the dataset.

[0042] The data is divided into training and testing sets for training the object detection network and calculating its quantitative evaluation metrics. To implement interpretability methods, representative samples need to be identified from the collected data; interpretations generated based on these samples can then be used to explain the decision-making process of the object detection network.

[0043] Step S3: Train the selected object detection network on the training set.

[0044] Step S3.1: Network selection: Select an object detection network suitable for the specific task, such as Faster R-CNN, YOLO-V3, SSD, etc.

[0045] Step S3.2: Network Initialization: Initialize the network by assigning initial values ​​to each parameter using parameter initialization methods. Step S3.3: Loss Function Selection: Select a suitable loss function to measure the difference between the model's predicted values ​​and the true values, and then train the network.

[0046] Step S3.4: Network Training: Train the object detection network using the training dataset. Forward and backward propagation are required during the training process.

[0047] Step S3.5: Parameter tuning: Adjust each hyperparameter based on the network training results and the value of the loss function to optimize the model's performance.

[0048] Step S4: Evaluate the object detection network using the test set and calculate quantitative evaluation metrics. Available metrics include accuracy, recall, precision, F1 score, etc. (See attached image) Figure 2 As shown.

[0049] Step S4.1: Model loading.

[0050] Load the parameters of the pre-trained object detection network model.

[0051] Step S4.2: Run the target detection network.

[0052] The trained object detection network is used to perform inference on the test set, calculate the position and category information of the regression boxes, and save the detection results.

[0053] Step S4.3: Calculate the evaluation indicators.

[0054] Based on the inference results and the labeled information in the dataset, calculate evaluation metrics such as accuracy, recall, and precision. Step S4.4: Performance Analysis.

[0055] Analyze the values ​​and distribution of evaluation metrics to understand the model's performance and limitations on the test dataset.

[0056] Step S4.5: Visualize the results.

[0057] The evaluation results are presented visually in the form of charts, reports, etc., so that others can understand and use them.

[0058] Step S5: Propose an interpretability method based on perturbation and class activation mapping to interpret the target detection network, analyze the interpretability of the target detection network, and calculate the interpretability score.

[0059] Step S5.1: Propose an interpretability method for object detection networks based on perturbation and class activation mapping, and present the interpretation results using saliency maps. (See attached image.) Figure 3 As shown.

[0060] Step S5.1.1: Generate a feature activation map mask.

[0061] A feature activation image mask is generated from the last convolutional layer of the object detection network. First, the output of each channel of the last convolutional layer is extracted. Then, the value of each channel is binarized, and the feature activation image f is generated. i C Convert to binary mask Its expression is as follows:

[0062]

[0063] In the formula, C represents the category, α is the threshold value (0,1), and i represents the i-th layer. Then, the binary mask is applied. The feature activation mask is obtained by applying bilinear interpolation and upsampling. Finally, the feature activation mask The feature activation image mask is obtained by performing a Hadamard product with the input image I. Its expression is as follows:

[0064]

[0065] Step S5.1.2: Calculate feature importance weights using IoU and similarity differences.

[0066] Activation map mask for all features The bounding box (bbox) and classification probability prediction score are obtained by feeding the data into the object detection network. These two parts together form the feature activation map mask. The predicted score vector is then calculated. Next, the IoU value between the anchor box of each feature activation image mask and the anchor box of the original input image is calculated. IoU is the overlap rate between the generated candidate box and the original label box. A high overlap rate between the feature activation image mask and the predicted class results in a higher correlation value; conversely, a low overlap rate results in a lower correlation value. The predicted vector of the original input image and the predicted vector of the i-th feature activation image mask are then compared. The IoU value between them is measured by the following formula:

[0067]

[0068] After calculating the IoU value, a similarity difference metric is calculated between the feature activation image mask and the original input image. These two vectors give the relevance of the feature activation image mask relative to the original input image. The intuition behind calculating the relevance of feature maps is to measure how the prediction changes when the features are unknown, i.e., the similarity difference between the predicted score sequences. The similarity difference metric is given by the following formula:

[0069]

[0070] Step S5.1.3: Calculate the feature importance weights and obtain the visual interpretation of the prediction.

[0071] The feature importance weight of each pixel can be obtained by taking the dot product of the IoU value and the similarity difference, as shown in the following formula:

[0072]

[0073] In the formula, W i C This represents the feature importance weight of each pixel. The feature activation mask is weighted and summed with its corresponding feature importance weight, as shown in the formula below:

[0074]

[0075] The visual interpretation diagram is represented in the form of a heatmap (saliency map), denoted as S. c .

[0076] Step S5.2: Analyze the interpretability results of the object detection network and calculate the interpretability score, as shown in the appendix. Figure 4 As shown.

[0077] The obtained visual interpretation map is quantized to obtain an interpretability score. In this embodiment, a deletion game is used to quantize the visual interpretation map. For each visual interpretation map, the k most important pixels are removed sequentially according to their given pixel feature importance scores, and then fed into the object detection network to calculate the change in the predicted score vector. The area enclosed by the polyline and the x-axis is calculated; the smaller the area, the better the interpretability method and the higher the interpretability score. The addition process is the reverse.

[0078] Step S6: Propose a visualization technique based on nonnegative matrix factorization and gradient descent to interpret the target detection network and analyze the semantics of the network structure in the target detection network.

[0079] Step S6.1: Build a custom network.

[0080] Common backpropagation-based object detection networks such as Faster R-CNN and YOLO-V3 contain non-linear structures such as activation functions, which can easily lead to problems such as gradient vanishing when using optimization algorithms. Therefore, linear layers are needed to replace the original activation function layers.

[0081] Step S6.2: Visualize the parameters of the object detection network.

[0082] The parameters visualized in this invention include the parameters of the convolution kernel and the parameters of the expanded weights.

[0083] The expansion weights can be understood as the magnitude of the interaction between feature maps. Suppose there are adjacent convolutional layers l1 and l2, whose generated feature maps are denoted as fm1 and fm2, respectively, in the form [w1, h1, channels1] and [w2, h2, channels2]. Taking the derivative of the mean of a certain channel of fm2 with respect to a certain channel of fm1, the final derivative is the expansion weight, denoted as W_l1_l2, in the form [width, height, channels1, channels2]. W_l1_l2[:,:, m, n] represents the interaction between the m-th convolutional kernel of l1 and the n-th convolutional kernel of l2.

[0084] This invention reduces the convolution kernel parameters and expanded weight parameters to 3 dimensions through Non-negative Matrix Factorization (NMF), and presents them in the form of images.

[0085] Step S6.3: Visualize the semantics learned by the structure in the object detection network.

[0086] In this invention, the objects that can be visualized can be the semantics learned by the convolution kernel, the semantics learned by the group of convolution kernels, the semantics learned by the convolutional layer, and the semantics learned by the nodes of the fully connected layer.

[0087] Step S6.3.1: Select the target structure for visualization.

[0088] A single convolutional kernel is the smallest unit of convolution operation in a convolutional neural network, often representing part or all of a semantic concept. A group of convolutional kernels, consisting of multiple kernels belonging to the same convolutional layer, can represent a more complete semantic concept. Convolutional layers contain a large number of convolutional kernels, and the resulting visualizations often contain complex and diverse semantics. Fully connected layer nodes are the foundation of classification, usually following convolutional layers, and decouple the semantics within the convolutional layers.

[0089] Step S6.3.2: Select a regularization technique.

[0090] Regularization techniques are used to personalize images synthesized using gradient descent. This invention proposes four regularization terms: L1 loss, L2 loss, total variance, and similarity penalty, to shift the synthesized image in the desired direction. L1 loss improves the contrast of the synthesized image, while L2 loss tends to produce synthesized images with smaller pixel values. The addition of a total variance regularization term reduces high-frequency noise in the synthesized image, and the similarity penalty ensures the synthesized image is similar to pre-set sample images.

[0091] Step S6.3.3: Run the gradient descent algorithm.

[0092] Using an optimization algorithm based on gradient descent, images that maximize the output of the target structure are generated iteratively starting from random input.

[0093] Step S7: Analyze the evaluation results.

[0094] The evaluation results of quantitative metrics and interpretability methods are comprehensively analyzed. The results of interpretability methods can be used to reveal the decision-making basis of the model and the semantics learned by neurons, helping users understand the predictions of object detection networks. For example, heatmaps generated by interpretability methods can reveal the magnitude of the contribution of features in an image to the prediction; observing the visualization results of neurons can help determine the semantics corresponding to the neurons (such as high- and low-frequency probes, black-and-white probes, curve probes, etc.).

[0095] As described above, the embodiments of the present invention explain the internal structure and decision-making basis of the target detection algorithm based on deep neural networks through interpretability methods, and construct an evaluation method for the target detection network by combining quantitative evaluation indicators and interpretability methods.

[0096] Based on the quantitative evaluation indicators and corresponding evaluation process provided by this invention, high-quality object detection models can be trained, and their performance indicators in practical tasks can be predicted. Based on the interpretability method provided by this invention, the knowledge and decision-making basis contained in deep neural network-based object detection models can be visualized, improving users' confidence in the object detection models.

[0097] Any obvious modifications, equivalent substitutions, or further optimizations made to the above details without departing from the basic technical concept and spirit of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A target detection network evaluation method based on quantification evaluation and interpretability method, characterized in that, The performance of the target detection algorithm based on the deep neural network is evaluated from different angles by integrating quantitative evaluation and explainability methods, the explainability of the target detection network is improved, the credibility of the model prediction result is enhanced, and the evaluation can be customized and adjusted according to different application scenarios and requirements to realize more accurate evaluation; the evaluation method comprises the following steps: Step S1: determining evaluation targets and indexes, and clearly defining the target detection network to be evaluated and related evaluation indexes; Step S2: data acquisition and preprocessing: collecting pictures and preprocessing to construct a training set for training the model, a test set for evaluating the performance of the model, and a sample for generating an explanation; Step S3: training the selected target detection network on the training set; Step S4: evaluating the target detection network using the test set, calculating the quantitative evaluation indexes, and the available indexes include accuracy, recall rate, average precision, and FPS; Step S5: proposing an explainability method based on perturbation and class activation mapping to explain the target detection network, analyzing the explainability of the target detection network, and calculating the explainability score; the explainability method based on perturbation and class activation mapping for the target detection network is proposed, and the explanation result is displayed in the form of a saliency map; a feature activation image mask is generated from the last convolution layer of the target detection network, the output of each channel of the last convolution layer is extracted, then the value of each channel is binarized, the feature activation map is converted into a binary mask, and the feature activation mask and the input image are multiplied to obtain the feature activation image mask; the IoU value between the anchor box of each feature activation image mask and the anchor box of the original input image is calculated, the IoU is the overlap rate of the generated candidate box and the original labeled box; the similarity difference measure between the feature activation image mask and the original input image is calculated, the IoU value and the similarity difference are dot multiplied to obtain the feature importance weight of each pixel, the feature activation mask and the corresponding feature importance are weighted and summed to obtain a visual explanation graph in the form of a heat map; the explainability score is calculated: the pixels with high scores are deleted in turn by using the deletion game, and the area surrounded by the probability score decline curve and the coordinate axis is calculated; Step S6: proposing a visualization technology based on non-negative matrix factorization and gradient descent method to explain the target detection network, and analyzing the semantics of the network structure in the target detection network; building a custom network: using a custom network structure, the difference between the custom network structure and the original target detection network is that the activation function layer is replaced by a linear layer; visualizing the parameters of the target detection network: visualizing the convolution kernel parameters and the extended weight parameters of the target detection network by using the non-negative matrix factorization method; visualizing the semantics learned by the structure in the target detection network: selecting the network structure to be visualized, and using an optimization algorithm based on the gradient descent method and the regularization technique to iteratively generate pictures that can maximize the output of the network structure from random input; Step S7: The evaluation results are analyzed and displayed.

2. The target detection network evaluation method based on quantification evaluation and interpretability method according to claim 1, characterized in that, In the step S1, the specific steps are as follows: The target network to be evaluated and the evaluation indicators are determined. The evaluation indicators are specific quantitative indicators used to measure whether the evaluation target is achieved.

3. The target detection network evaluation method based on quantification evaluation and interpretability method according to claim 1, characterized in that, In step S2, the specific steps are as follows: Step S2.1: Determine the content and size of the data set: determine the type and number of pictures to be collected based on task requirements and characteristics; Step S2.2: Data collection and preprocessing: collect data suitable for target detection network training and testing, and perform appropriate preprocessing; Step S2.3: Label the pictures and save them: label the regression boxes for the pictures and save them; Step S2.4: Build the data set: divide the collected data into training and testing sets and select samples for explainability methods.

4. The target detection network evaluation method based on quantification evaluation and interpretability method according to claim 1, characterized in that, In step S3, the specific steps are as follows: Step S3.1: Network selection: select a target detection network suitable for a specific task, including Faster RCNN, YOLO-V3, and SSD; Step S3.2: Network initialization: initialize the network, i.e., use the parameter initialization method to assign initial values to the parameters of the network; Step S3.3: Loss function selection: select an appropriate loss function to measure the difference between the model's predicted value and the true value; Step S3.4: Network training: use the training set to train the target detection network. Forward propagation and backpropagation are required during training; Step S3.5: Parameter tuning: adjust the hyperparameters based on the results of network training and the value of the loss function to optimize the performance of the model.

5. The target detection network evaluation method based on quantification evaluation and interpretability method according to claim 1, characterized in that, In step S4, the specific steps are as follows: Step S4.1: Model loading: load the trained target detection network model parameters; Step S4.2: Run the target detection network: use the trained target detection network to perform inference on the test set, calculate the position and category information of the regression box; Step S4.3: Calculate evaluation indicators: calculate evaluation indicators based on the inference results and the annotation information in the data set, including accuracy, recall, and precision; Step S4.4: Performance analysis: analyze the values and distribution of the evaluation indicators to determine the performance and limitations of the model on the test set; Step S4.5: Result visualization: visualize the evaluation results in the form of charts and reports to facilitate understanding and use by others.

6. The method of claim 1, wherein the method of evaluating a target detection network based on quantification evaluation and interpretability comprises, In step S7, the evaluation results are analyzed, including: Integrate the evaluation results of quantitative and explainability methods and generate interpretation in the form of charts and heat maps to facilitate user understanding and use.

Images