Reverse distillation image anomaly detection method and device based on feature regulation and control, and medium

Abstract

The invention provides a feature regulation-based anti-distillation image anomaly detection method and device and a medium, and belongs to the technical field of artificial intelligence. Comprising the following steps: constructing a redistillation network; normal images are collected for natural synthesis abnormal operation; enabling the synthesized image to sequentially pass through a teacher model and a bottleneck layer to obtain compact features; recovering features from the compact features through a student model, and generating normal pixel features and abnormal pixel features in combination with a feature mask and a recovery module; the difference between the normal pixel features and the abnormal pixel features relative to the teacher model output features is calculated, the normal pixel feature distance and the abnormal pixel feature distance are calculated, and a back distillation network is trained in combination with difference loss; and inputting a to-be-detected image into the trained back distillation network, calculating a cosine similarity difference graph of output features of the teacher model and the student model, fusing to generate an abnormal scoring graph, and positioning an abnormal region. According to the method, the industrial product surface anomaly positioning reliability is improved, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The present invention relates to an anti-distillation image anomaly detection method, device and medium based on feature regulation, belonging to the field of artificial intelligence technology. Background Art

[0002] In recent years, anomaly detection methods based on knowledge distillation have received extensive attention due to their advantages in feature expression and difference learning. Such methods (such as STPM, MRKD, etc.) construct a teacher-student model framework, use the teacher model to extract normal sample features and guide the student model to learn, and then locate the anomaly region through feature differences. Specifically, the STPM method adopts a feature pyramid matching strategy to transfer normal pattern knowledge by aligning multi-scale features layer by layer; the MRKD method optimizes the discrimination ability of the feature space by introducing a mask technology and a dual-teacher model architecture. However, the existing technologies still have the following key defects: Feature distribution overlap problem: Since the distributions of normal and abnormal features in the low-dimensional space highly overlap, it is difficult for the student model to accurately capture subtle differences, resulting in a significant decline in the reliability of anomaly localization, especially with a high false detection rate in complex texture or weak anomaly scenarios. Multi-scale noise interference problem: During the feature pyramid or multi-scale propagation process, low-level features (such as local texture, illumination noise) extracted by the shallow network will interfere with the expression of high-level semantic features through cascade operations, reducing the sensitivity of the model to anomaly regions. Computational efficiency bottleneck: Existing methods often adopt complex architectures (such as multi-teacher models, dense feature interaction modules) to improve performance, resulting in an exponential growth of computational complexity and making it difficult to meet the stringent real-time requirements of industrial detection scenarios.

[0003] The above defects pose severe challenges to existing methods in fields that emphasize both accuracy and efficiency. Summary of the Invention

[0004] The object of the present invention is to provide an anti-distillation image anomaly detection method, device and medium based on feature regulation, which improves the reliability of anomaly localization on the surface of industrial products and enhances the semantic robustness of the model.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: An anti-distillation image anomaly detection method based on feature regulation, comprising the following steps: Construct an anti-distillation network, including a pre-trained teacher model, a bottleneck layer, a student model, a feature mask and restoration module, and a collaborative difference optimization and loss calculation module. The bottleneck layer introduces a feature selection module, and the student model is a decoding network with a structure symmetric to that of the teacher model in reverse; Collect normal images for natural synthesis with abnormal operations to generate synthetic images containing normal pixels and abnormal pixels; Input the synthetic image into the teacher model to extract multi-scale features, dynamically adjust the weights of different-scale features through the feature selection module of the bottleneck layer, and compress to obtain compact features; Recover features from the compact features through the student model, and generate normal pixel features and abnormal pixel features in combination with the feature mask and the recovery module; Calculate the differences between the normal pixel features and the abnormal pixel features relative to the output features of the teacher model, calculate the normal pixel feature distance and the abnormal pixel feature distance through the collaborative difference optimization and loss calculation module, and train the anti-distillation network in combination with the difference loss; Input the image to be tested into the trained anti-distillation network, extract multi-scale features through the teacher model, and input them into the student model to generate recovered features after being processed by the bottleneck layer; Calculate the cosine similarity difference map of the output features of the teacher model and the student model, and fuse to generate an abnormal score map and locate the abnormal area.

[0006] Preferably, the teacher model is a WideResNet50 model.

[0007] Preferably, the feature selection module dynamically adjusts the weights of different-scale features and compresses to obtain compact features, specifically including: Adopt multi-scale feature fusion to fuse the output features of the teacher model into one feature, perform global average pooling on this feature and then generate channel weights through one-dimensional convolution; Multiply the channel weights by the output features of the teacher module to obtain weighted features; Compress the weighted features through a type of embedding to generate compact features.

[0008] Preferably, the formula for generating channel weights by one-dimensional convolution is: , where: represents the element of the compact feature, represents of a set of adjacent channels, represents the weight parameter of the th adjacent channel in the one-dimensional convolution kernel, represents the attention weight of the th channel, represents the Sigmoid function, represents the th channel's th adjacent channel's feature value.

[0009] Preferably, the normal pixel feature distance is the deviation of the cosine similarity between the normal feature output by the student model and the normal feature input by the teacher model , and the formula is as follows: where, represents the normal feature output by the student model, represents the normal feature input by the teacher model, represents the transpose, represents the set of normal pixels, is the number of the encoding block, represents the normal feature; The abnormal pixel feature distance is the deviation of the cosine similarity between the abnormal feature output by the student model and the abnormal feature input by the teacher model , and the formula is as follows: , where, represents the abnormal feature input by the teacher model, represents the abnormal feature output by the student model, represents the set of abnormal pixels, represents the abnormal feature.

[0010] Preferably, the weighted loss function of the collaborative difference optimization module is: , where, represents the weight of the th normal pixel in the th encoding block, represents the cosine distance of the th normal pixel in the th encoding block, represents the weight of the th abnormal pixel in the th encoding block, represents the cosine distance of the th abnormal pixel in the th encoding block, represents the number of normal pixels in the th encoding block, represents the number of abnormal pixels in the th encoding block.

[0011] Preferably, the weight of the normal pixel, and the formula is as follows: , , where, represents the mean value of the normal feature distance, is a hyperparameter; The abnormal pixel weight is given by the following formula: , , where, represents the mean of the normal feature distances.

[0012] Preferably, the abnormal score map is generated through the following steps: Perform bilinear upsampling on the cosine similarity difference map; generate the final score map through pixel-by-pixel multiplication and accumulation.

[0013] A feature-regulation-based anti-distillation image anomaly detection device includes a processor and a memory storing program instructions. The processor is configured to execute the feature-regulation-based anti-distillation image anomaly detection method when running the program instructions.

[0014] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the feature-regulation-based anti-distillation image anomaly detection method.

[0015] The advantages of the present invention are as follows: Through the collaborative difference optimization module, the present invention can effectively distinguish normal pixels from abnormal pixels through the masking technique. During training, the feature distances of normal pixels are reduced to maintain consistency, while the feature distances of abnormal pixels are increased to clearly distinguish them. The feature selection module is proposed to learn the correlation of features at different scales, reduce the influence of abnormal information on the subsequent inference process, and improve the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention.

[0017] Figure 1 is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 is a schematic diagram of the feature processing flow of the present invention.

[0019] Figure 3 is a schematic diagram of the feature selection module of the present invention.

[0020] Figure 4 is a comparison chart of the detection effects of different methods in Simulation Experiment 1.

[0021] Figure 5 is a comparison chart of the detection effects of different methods in Simulation Experiment 2.

[0022] Figure 6Collect normal images for the embodiments.

[0023] Figure 7 Naturally synthesize abnormal images. Detailed implementation manners

[0024] 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. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] Embodiment 1 As Figure 1 shown, an anti-distillation image anomaly detection method based on feature regulation can effectively distinguish normal pixels from abnormal pixels through a collaborative difference optimization module using a masking technique. During training, the feature distance of normal pixels is reduced to maintain consistency, while the feature distance of abnormal pixels is increased to clearly distinguish them. A feature selection module is proposed to learn the correlation of features at different scales, reduce the influence of abnormal information on the subsequent inference process, and improve the generalization ability of the model.

[0026] Specifically, it is applied to the identification of surface defects of industrial products and includes a training stage and an inference stage.

[0027] Training stage: S1: Construct an anti-distillation network, including a pre-trained teacher model, a bottleneck layer, a student model, a feature mask and recovery module, and a collaborative difference optimization and loss calculation module. The bottleneck layer introduces a feature selection module, and the student model is a decoding network with a structure symmetric to the teacher model in the reverse direction.

[0028] As a refinement of the above embodiment, the teacher model is a pre-trained WideResNet50, and the student model is a decoding network with a structure symmetric to the teacher model in the reverse direction.

[0029] S2: Collect normal images for natural synthesis of abnormal operations to generate synthetic images containing normal pixels and abnormal pixels. The images are surface defect images of industrial products.

[0030] For example: Normal image collection: Capture the surface image of a defect-free screw through a high-resolution industrial camera to ensure that the thread is complete and the head has no cracks, as Figure 6 .

[0031] Natural synthesis of abnormality (NSA): Use Poisson image editing technology to seamlessly embed abnormal patches into normal screw images to generate synthetic training samples, Figure 7 showing an abnormal area with an obvious fracture in the screw head.

[0032] As a refinement of the above embodiment, the natural synthesis anomaly operation (NSA) generates a synthetic image containing normal pixels through the following formula and abnormal pixels as follows : , where is the gradient operator, is the target image, is a randomly selected rectangular patch

[0033] S3: Input the synthetic image into the teacher model to extract multi-scale features ( ) and dynamically adjust the weights of different-scale features through the feature selection module of the bottleneck layer, and compress to obtain compact features

[0034] As a refinement of the above embodiment, in order to reduce the introduction of abnormal information and dynamically adjust the information of different-scale features, a feature selection module is introduced in the bottleneck layer. The specific structure is as Figure 3 shown. The aggregated features obtained by global average pooling generate weights through a fast one-dimensional convolution of size

[0035] Specifically, it includes First, use multi-scale feature fusion MFF (Multi-scale feature fusion) to fuse the , , output by the teacher network to obtain ; Adjust it to through a global average pooling, and then obtain the channel weight of the th channel after one-dimensional convolution , where represents the element of the compact feature, represents of the adjacent channel set, represents the weight parameter of the th adjacent channel in the one-dimensional convolution kernel, represents the attention weight of the th channel. After being normalized by the Sigmoid function, the range is , which is used to represent the importance of this channel, represents the Sigmoid function to ensure the interpretability of the attention weight​ representing the eigenvalue of the th adjacent channel of the th channel; a weight parameter corresponds to the eigenvalue of an adjacent channel. Specifically, in the channel dimension, centered on the th channel, the eigenvalues of its left and right neighbors (including itself) are taken. This strategy can be implemented by a fast one-dimensional convolution with a kernel size of , where represents one-dimensional convolution, and the channel weight is multiplied by the original feature to obtain the weighted feature , and the formula is as follows: , The feature is transmitted to a one-class embedding (OCE) for information compression to obtain ; this step effectively reduces the propagation of abnormal perturbations and improves the robustness of the student model in capturing normal patterns.

[0036] S4: The student model recovers the feature from the compact feature, and combines the feature mask and the recovery module to generate normal pixel features and abnormal pixel features.

[0037] As a refinement of the above embodiment, the compact feature output by the bottleneck layer is generated after the multi-scale features extracted by the teacher model are compressed and filtered by the bottleneck layer.

[0038] The student model adopts a structure that is symmetrically opposite to the teacher model (the teacher model is an encoder, and the student model is a decoder), and recovers the feature from through reverse operations.

[0039] The feature mask and the recovery module further include a mask generation module , which dynamically generates a mask to suppress the interference of abnormal regions or enhance the details of normal regions. The weights of the recovered features are adjusted through lightweight network layers (such as convolutional layers), and the parameters are randomly initialized to improve flexibility.

[0040] The features output by the feature mask and the recovery module include: the normal pixel features recovered by the student model, and the abnormal pixel features recovered by the student model.

[0041] The mathematical expression is as follows: , 。

[0042] Through reverse structure and mask generation, the student model learns to align abnormal features with normal features while retaining the ability to distinguish details.

[0043] S5: Calculate the differences between the normal pixel features and the abnormal pixel features with respect to the output features of the teacher model. Calculate the normal pixel feature distance and the abnormal pixel feature distance through the collaborative difference optimization and loss calculation module, and train the anti-distillation network by combining the difference loss.

[0044] As a refinement of the above embodiment, the feature distance calculation method is as follows: Normal pixel distance , that is, in the th encoding block, the feature distance of the normal pixel set ranges from , and the larger the value, the greater the difference between the two. It is obtained by deviating from the cosine similarity between the normal feature restored by the student and the original teacher feature : , where represents the normal feature output by the student model, represents the normal feature input by the teacher model, represents transpose, represents the normal pixel set, is the number of the encoding block, corresponding to a certain level of the teacher model and the student model (such as different convolutional blocks in WideResNet50). By calculating the feature distance layer by layer, multi-scale anomaly detection is achieved. represents the normal feature.

[0045] is the inner product of the teacher feature and the student feature , used to measure the directional similarity between the two.

[0046] Abnormal pixel distance , that is, in the th encoding block, the feature distance of the abnormal pixel set ranges from , and the larger the value, the greater the difference between the two. It is obtained by deviating from the cosine similarity between the abnormal feature restored by the student and the original teacher feature : , where represents the abnormal feature input by the teacher model, Represents the abnormal features output by the student model, Represents the set of abnormal pixels, Represents the abnormal features.

[0047] The weighted loss function of the collaborative difference optimization module is: , Among them, Represents the th normal pixel weight in the th coding block, Represents the th normal pixel cosine distance in the th coding block, Represents the th abnormal pixel weight in the th coding block, Represents the th abnormal pixel cosine distance in the th coding block, Represents the th number of normal pixels in the coding block, Represents the th number of abnormal pixels in the coding block.

[0048] Molecular logic: Normal term: Minimize the normal pixel distance to make the features recovered by the student model close to those of the teacher model.

[0049] Abnormal term: Maximize the abnormal pixel distance to enhance the distinguishability of abnormal features.

[0050] Denominator function: Normalize the weights to avoid optimization bias caused by uneven sample numbers.

[0051] The dynamic weight adjustment of the loss function is as follows: Mean calculation, calculate the means of the normal and abnormal feature distances respectively: , , Weight allocation: Normal pixel weight: , the larger it is than the mean, the higher the weight, forcing the model to preferentially optimize difficult samples.

[0052] Abnormal pixel weight: , the smaller it is than the mean, the higher the weight, avoiding the model from over-recovering abnormal features.

[0053] Among them, is a hyperparameter for the adjustment effect. The exponent of normal pixels is , and the exponent of abnormal pixels is - . The optimal value is 2. When it increases from 1.5 to 2, the performance gradually improves. When it is greater than 2, the performance begins to decline.

[0054] Inference stage: S6: Input the image to be tested into the trained anti-distillation network, extract multi-scale features through the teacher model, and input them into the student model to generate restored features after being processed by the bottleneck layer.

[0055] As a refinement of the above embodiment, the cosine similarity difference map : , where are the spatial coordinates of the feature map, corresponding to the position of a local area in the original image. For example: in the image, represents the central area.

[0056] represents the feature vector extracted by the teacher model in the th encoding block, and the position is .

[0057] represents the feature vector restored by the student model in the th decoding block, and the position is .

[0058] S7: Calculate the cosine similarity difference map of the output features of the teacher model and the student model, fuse and generate an anomaly score map and locate the anomaly area.

[0059] As a refinement of the above embodiment, the anomaly score map is generated through the following steps: Perform bilinear upsampling on the cosine similarity difference map; generate the final score map through pixel-by-pixel multiplication and accumulation , and the formula is as follows: .

[0060] Embodiment 2 Simulation experiment 1, which simulates the present invention on the MVTec AD dataset.

[0061] Simulation conditions: Conducted under Pytorch 1.12.1 and python 3.8 software.

[0062] Such as Figure 4As shown in the figure, in the MVTecAD industrial dataset (covering multiple types of product surface defects), the feature-controlled reverse distillation anomaly detection method is compared with mainstream methods such as MRKD, PatchCore, and PaDiM. The experimental results show that the feature-controlled reverse distillation anomaly detection method exhibits more accurate positioning capabilities in complex texture scenes (such as carpet fiber breakage) and minor defects (such as capsule surface scratches). Among the comparison methods, MRKD has a high false detection rate for abnormal areas due to the lack of feature control mechanism; although PatchCore can capture local features, it lacks the fusion of multi-scale information, resulting in missed detection of minor defects. Through the collaborative difference optimization module (CDO), the feature-controlled reverse distillation anomaly detection method significantly distinguishes the normal and abnormal feature distributions. Its abnormal score distribution map shows that normal samples are concentrated in low-score areas, and abnormal samples are significantly offset and clearly spaced. As can be seen from the comparison of the attached figure, the positioning results of the feature-controlled reverse distillation anomaly detection method are more consistent with the real mask, especially in the "cable" category with complex textures, the false detection area is significantly reduced.

[0063] Example 3 Simulation experiment 2 is a simulation of the present invention on the BTAD dataset.

[0064] Simulation conditions: carried out under Pytorch1.12.1, python3.8 software.

[0065] like Figure 5 As shown in the figure, based on the BTAD real industrial scene dataset (including interference such as illumination changes and occlusion), the feature-controlled reverse distillation anomaly detection method is compared with VT-ADL, AE (MSE+SSIM), RD4AD and other methods. The results show that the feature-controlled reverse distillation anomaly detection method can still stably detect anomalies in challenging scenarios such as reflections on metal parts and low-contrast cracks on plastic surfaces, while the comparison methods have positioning deviations or missed detections due to the single feature extraction or insufficient dynamic adjustment capabilities. For example, VT-ADL mistakenly identifies background noise as a defect in a scene with stains on the surface of a screw with severe occlusion; the feature-controlled reverse distillation anomaly detection method dynamically fuses multi-scale information through a feature selection module (SF), effectively suppresses interference and captures subtle abnormal patterns. In the attached figure, the target boundary accuracy of the feature-controlled reverse distillation anomaly detection method exceeds the benchmark method in the occlusion scene of "category 2", verifying its anti-interference ability.

[0066] An embodiment of the present disclosure also provides an anti-distillation image anomaly detection device based on feature regulation, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. Among them, the processor, the communication interface, and the memory can complete mutual communication through the bus. The communication interface can be used for information transmission. The processor can call the logical instructions in the memory to execute the anti-distillation image anomaly detection method based on feature regulation in the above embodiment.

[0067] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as an independent product, they can be stored in a computer-readable storage medium.

[0068] As a computer-readable storage medium, the memory can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of the present disclosure. The processor executes functional applications and data processing by running the program instructions / modules stored in the memory, that is, implements the anti-distillation image anomaly detection method based on feature regulation in the above embodiment.

[0069] The memory may include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0070] An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, and the computer-executable instructions are set to execute the anti-distillation image anomaly detection method based on feature regulation described above.

[0071] The above-mentioned computer-readable storage medium can be a transient computer-readable storage medium or a non-transient computer-readable storage medium.

[0072] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature-regulated anti-distillation image anomaly detection method, characterized in that, It includes the following steps: Construct an anti-distillation network, including a pre-trained teacher model, a bottleneck layer, a student model, a feature mask and recovery module, and a collaborative difference optimization and loss calculation module. The bottleneck layer introduces a feature selection module, and the student model is a decoding network with a structure symmetric to that of the teacher model in the reverse direction; Collect normal images and perform natural synthesis and anomaly operations to generate synthetic images containing normal pixels and abnormal pixels; Input the synthetic images into the teacher model to extract multi-scale features, dynamically adjust the weights of different-scale features through the feature selection module of the bottleneck layer, and compress them to obtain compact features; Restore the features from the compact features through the student model, and generate normal pixel features and abnormal pixel features in combination with the feature mask and recovery module; Calculate the differences between the normal pixel features and the abnormal pixel features with respect to the output features of the teacher model, calculate the normal pixel feature distance and the abnormal pixel feature distance through the collaborative difference optimization and loss calculation module, and train the anti-distillation network in combination with the difference loss; Input the image to be tested into the trained anti-distillation network, extract multi-scale features through the teacher model, and input them into the student model to generate restored features after being processed by the bottleneck layer; Calculate the cosine similarity difference map of the output features of the teacher model and the student model, and fuse them to generate an anomaly score map and locate the anomaly region.

2. The method for anomaly detection of anti-distillation images based on feature regulation according to claim 1, wherein, The teacher model is a WideResNet50 model.

3. The method for anomaly detection of anti-distillation images based on feature regulation according to claim 1, wherein, The feature selection module dynamically adjusts the weights of different-scale features and compresses them to obtain compact features, specifically including: Adopt multi-scale feature fusion to fuse the output features of the teacher model into one feature, perform global average pooling on this feature and then generate channel weights through one-dimensional convolution; Multiply the channel weights by the output features of the teacher module to obtain weighted features; Compress the weighted features through a type of embedding to generate compact features.

4. The method for anti-distillation image anomaly detection based on feature regulation according to claim 3, wherein The formula for generating channel weights by one-dimensional convolution is: ， Wherein: An element representing a compact feature, represents a set of adjacent channels, represents the weight parameter of the th adjacent channels in a one-dimensional convolutional kernel, represents the attention weight of the th channel, represents the eigenvalue of the th adjacent channel of the th channel.

5. The method for anomaly detection of anti-distillation images based on feature regulation according to claim 1, characterized in that The normal pixel feature distance is the deviation of the cosine similarity between the normal features output by the student model and the normal features input by the teacher model , and the formula is as follows: Among them, represents the normal features output by the student model, represents the normal features input to the teacher model, represents transpose, represents the set of normal pixels, is the number of the encoding block, represents normal features; The abnormal pixel feature distance is the deviation of the cosine similarity between the abnormal feature output by the student model and the abnormal feature input by the teacher model , and the formula is as follows: ， Among them, represents the abnormal features input to the teacher model, represents the abnormal features output by the student model, represents the set of abnormal pixels, represents the abnormal features.

6. The feature-regulation-based anti-distillation image anomaly detection method according to claim 1, wherein, The weighted loss function of the collaborative difference optimization module is: ， Among them, represents the th normal pixel weight in the th coding block, represents the th normal pixel cosine distance in the th coding block, represents the th abnormal pixel weight in the th coding block, represents the th abnormal pixel cosine distance in the th coding block, represents the number of normal pixels in the th coding block, represents the number of abnormal pixels in the th coding block.

7. The method for anomaly detection of anti-distillation images based on feature regulation according to claim 6, characterized in that, The weight of the normal pixel, the formula is as follows: ， ， Among them, represents the mean of the normal feature distances, is a hyperparameter; The weight of the abnormal pixel, the formula is as follows: ， ， Among them, represents the mean of the normal feature distances.

8. The method for anomaly detection of anti-distillation images based on feature regulation according to claim 1, wherein, The anomaly score map is generated through the following steps: Perform bilinear upsampling on the cosine similarity difference map; generate the final score map through pixel-by-pixel multiplication and accumulation.

9. An anti-distillation image anomaly detection device based on feature regulation, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the anti-distillation image anomaly detection method based on feature regulation as described in any one of claims 1-8 when running the program instructions.

10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and when the program is executed by the processor, it implements the anti-distillation image anomaly detection method based on feature regulation as described in any one of claims 1-8 above.

Images