Anomaly score map optimization method and system for industrial image anomaly detection

Abstract

Provided in the present invention are an anomaly score map optimization method and system for industrial image anomaly detection. Further optimization is performed on an anomaly score map that is extracted by an existing anomaly detection method based on unsupervised or self-supervised learning. A foreground noise suppression module and a background anomaly suppression module are designed. An anomaly score map to be optimized is obtained by means of an existing method. The anomaly score of a pseudo-anomaly in a foreground region is reduced by means of foreground noise suppression processing, so as to suppress the impact of the pseudo-anomaly in the foreground region during an anomaly detection process. An optimization mark related to the image foreground or object region is generated by means of the background anomaly suppression module, so that the anomaly score map focuses only on an anomaly region in the image foreground or object region, thereby suppressing the impact of a pseudo-anomaly in the background region. The present invention is independent of product categories, does not require additional training, is not limited to a specific anomaly detection method based on unsupervised or self-supervised learning, and can be used as a plug-in in a plug-and-play manner.

Description

An Anomaly Fraction Map Optimization Method and System for Industrial Image Anomaly Detection

[0001] Technical Field

[0002] This invention belongs to the field of image recognition technology, specifically relating to an anomaly score map optimization method and system for anomaly detection in industrial images. Background Technology

[0003] Artificial intelligence-based automatic anomaly detection methods have been widely applied in industrial fields. One common solution is based on unsupervised / self-supervised learning methods, which learn by modeling a large number of normal samples to obtain a distribution or representation of normal images. In practical applications, the difference between the sample to be detected and the modeled normal distribution or representation is calculated to obtain an anomaly score map for identifying / locating abnormal regions.

[0004] However, since no real anomalous regions / samples have been encountered during training, anomaly detection methods based on unsupervised or self-supervised learning often face significant limitations in identifying / localizing "pseudo-anomaly regions" in the samples to be detected—regions that differ from normal regions but are not truly anomalous. These pseudo-anomaly regions cannot be distinguished from genuine anomalous regions. Specifically: 1) Foreground or object regions in an image, there are often pseudo-anomaly regions caused by noise. These pseudo-anomaly regions also have high anomalous scores on the anomaly score map, thus affecting the final detection / localization results. 2) Some pseudo-anomaly regions in the image background also have high anomalous scores on the anomaly score map. However, the problem of anomaly detection / localization focuses on the foreground or objects in the image, and these pseudo-anomalies in the background also affect the final detection / localization results. Summary of the Invention

[0005] The purpose of this invention is to provide an anomaly score map optimization method and system for industrial image anomaly detection. For anomaly score maps extracted by existing anomaly detection methods based on unsupervised or self-supervised learning, this invention suppresses the abnormal scores of pseudo-anomaly regions in the foreground and background of the sample under test by suppressing foreground noise and background anomaly, respectively, thereby further optimizing the anomaly score map and improving the ability to identify / locate anomaly regions. Furthermore, this invention is not dependent on specific product categories, requires no additional training, and is not limited to any particular anomaly detection method based on unsupervised or self-supervised learning; it can be used plug-and-play as a plugin.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an anomaly score map optimization method for industrial image anomaly detection, comprising the following steps:

[0007] For the initial anomaly score map to be optimized, the influence of spurious anomalies in the foreground region during anomaly detection is suppressed to obtain an anomaly score map with suppressed foreground noise.

[0008] Background anomaly suppression is performed on the industrial image to be detected to obtain a marker map for distinguishing between background and foreground regions. Based on the marker map for distinguishing between background and foreground regions, outliers in the background region in the anomaly score map are removed to obtain an optimized anomaly score map.

[0009] Furthermore, obtain the initial anomaly score map to be optimized, including:

[0010] For the industrial image data to be detected, an initial anomaly score map to be optimized is obtained by using existing anomaly detection methods based on unsupervised or self-supervised learning.

[0011] The anomaly detection method based on unsupervised or self-supervised learning learns by modeling normal samples to obtain a model of the distribution or representation of normal images, and obtains an anomaly score map by calculating the difference between the sample to be detected and the modeled normal distribution or representation.

[0012] Furthermore, for the initial anomaly score map to be optimized, the influence of pseudo-anomalies in the foreground region during anomaly detection is suppressed to obtain an anomaly score map with suppressed foreground noise. This includes: smoothing the anomaly scores in local regions of the initial anomaly score map to be optimized by using a Gaussian kernel with set size and variance, reducing the anomaly scores in pseudo-anomaly regions on the foreground or objects, and obtaining an anomaly score map with suppressed foreground noise.

[0013] Furthermore, background anomaly suppression is performed on the industrial image to be detected to obtain a marker map for distinguishing between background and foreground regions. Based on the marker map, outliers located in the background region are removed from the anomaly score map to obtain an optimized anomaly score map, including:

[0014] Using pre-designed category-independent text cue words, the industrial image to be detected is input into a trained semantic segmentation model to obtain multi-labeled images related to the foreground cue words. and multi-marker graphs related to object cues In the multi-label map, the value of the relevant region is 1, and the value of the irrelevant region is 0. The above multi-label map is merged by taking the union to obtain the initial non-background region label, where the non-background region value is 1 and the background region value is 0.

[0015] The initial non-background area is processed as follows: mark areas that are not foreground or object are removed, hole noise areas are removed, and dilation is performed to obtain a mark map for distinguishing between background and foreground areas;

[0016] The marker map used to distinguish between background and foreground regions is multiplied element-wise with the anomaly score map that suppresses foreground noise to obtain the optimized anomaly score map.

[0017] Furthermore, based on the above-mentioned technical concept, this invention also provides an anomaly score map optimization system for industrial image anomaly detection, including a foreground noise suppression module and a background anomaly suppression module. The foreground noise suppression module is used to suppress the influence of pseudo-anomalies in the foreground region during anomaly detection for the initial anomaly score map to be optimized, thereby obtaining an anomaly score map with suppressed foreground noise. The background anomaly suppression module is used to perform background anomaly suppression on the industrial image to be detected to obtain a marker map for distinguishing between the background and foreground regions. Based on the marker map for distinguishing between the background and foreground regions, outliers located in the background region in the anomaly score map are removed to obtain the optimized anomaly score map.

[0018] Furthermore, the foreground noise suppression module takes as input the initial anomaly score map to be optimized. The output is an anomaly score map with foreground noise suppressed. The calculation is as follows:

[0019]

[0020] in, Let be a Gaussian filter function, with a Gaussian kernel size of . The variance is 5.

[0021] Furthermore, the background anomaly suppression module includes a non-background labeled region extraction unit and a background anomaly elimination unit. The non-background labeled region extraction unit includes a pre-trained semantic segmentation model. The industrial image to be detected is used as input to the semantic segmentation model. Pre-designed category-independent text prompts are combined with the image to be detected and input into the trained semantic segmentation model to obtain multi-labeled maps related to the foreground prompts. and multi-marker graphs related to object cues In the multi-label map, the value of the relevant region is 1, and the value of the irrelevant region is 0. The above multi-label map is merged by taking the union to obtain the initial non-background region label, where the non-background region value is 1 and the background region value is 0.

[0022] The initial non-background area is processed as follows: mark areas without foreground or object are removed, hole noise areas are removed, and a dilation operation is performed to obtain a marker map used to distinguish between background and foreground areas. ;

[0023]

[0024] The background anomaly removal unit uses the extracted marker map to distinguish between background and foreground regions. Anomaly score plot with suppressed foreground noise Perform element-wise multiplication to obtain the optimized anomaly score map. :

[0025] .

[0026] When the above-mentioned anomaly score map optimization method for industrial image anomaly detection is applied, the anomaly score map output by the anomaly detection method is optimized to obtain an optimized anomaly score map, which is then used as input to the industrial image anomaly recognition and localization method.

[0027] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the anomaly score map optimization method for industrial image anomaly detection described in the present invention when executing the executable program.

[0028] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, it can implement the anomaly score map optimization method for industrial image anomaly detection described in this invention.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] This invention provides an anomaly score map optimization method for industrial image anomaly detection. By suppressing foreground noise and background anomalies, the abnormal scores of pseudo-anomaly regions in the foreground and background of the sample to be detected can be suppressed in the anomaly score map. This solves the shortcomings of existing anomaly detection methods based on unsupervised or self-supervised learning and improves the ability to identify / locate anomaly regions.

[0031] The method provided by this invention is simple to use and can be plugged in as a plugin. It is not limited to a specific anomaly detection method based on unsupervised or self-supervised learning. It can effectively enhance the anomaly score map extracted by existing methods. The anomaly score map obtained by the upstream anomaly detection method can be directly used as input, and the output value can be used as input for the downstream anomaly localization or identification method.

[0032] The method provided by this invention requires no additional training process, is not dependent on specific product categories, and does not require separate processing of images of a single product category. It can be widely applied to real industrial scenarios with a wide variety of product categories.

[0033] Attached image content

[0034] Figure 1 is a schematic diagram of an implementable process of the present invention.

[0035] Figure 2 is a diagram of the architecture of the anomaly score map optimization method for industrial image anomaly detection proposed in this invention.

[0036] Figure 3 is a framework diagram of the anomaly score extraction module.

[0037] Figure 4 is a framework diagram of the foreground noise suppression module.

[0038] Figure 5 is a framework diagram of the non-background marked region extraction unit in the background anomaly suppression module.

[0039] Figure 6 is a framework diagram of the background anomaly elimination unit in the background anomaly suppression module.

[0040] Figure 7 shows the effect of the anomaly score map optimization method for industrial image anomaly detection proposed in this invention. Embodiments of the present invention

[0041] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0042] This invention proposes an anomaly score map optimization method for industrial image anomaly detection. By further optimizing the anomaly score map extracted by existing anomaly detection methods based on unsupervised or self-supervised learning, it improves the ability to identify / locate anomaly regions. As shown in Figure 2, to reduce the influence of pseudo-anomaly regions in the sample to be detected, this invention designs a foreground noise suppression module and a background anomaly suppression module. The anomaly score map to be optimized is obtained through an anomaly score map extraction module composed of existing anomaly detection methods based on unsupervised or self-supervised learning. Then, the foreground noise suppression module reduces the anomaly score of pseudo-anomalies in the foreground region, thereby suppressing the influence of pseudo-anomalies in the foreground region during anomaly detection. Finally, the background anomaly suppression module generates optimized labels related to the image foreground or object region, further optimizing the anomaly score map so that it only focuses on anomaly regions in the image foreground or object region, thereby suppressing the influence of pseudo-anomalies in the background region. This invention is not dependent on product category, requires no additional training, and is not limited to a specific anomaly detection method based on unsupervised or self-supervised learning; it can be used plug-and-play as a plugin.

[0043] As shown in Figure 1, an anomaly score map optimization method for industrial image anomaly detection specifically includes the following steps:

[0044] S1, Obtain the anomaly score map to be optimized.

[0045] As shown in Figure 3, for the industrial image data to be detected, the anomaly score map extraction module obtains an initial anomaly score map to be optimized, denoted as . The anomaly score map extraction module is used to implement existing anomaly detection methods based on unsupervised or self-supervised learning, but is not limited to a specific method. The anomaly detection method based on unsupervised or self-supervised learning is to model and learn from a large number of normal samples to obtain a model of the distribution or representation of normal images, and to obtain an anomaly score map by calculating the difference between the sample to be detected and the modeled normal distribution or representation for identifying and locating abnormal regions.

[0046] S2, after foreground noise suppression processing

[0047] For the initial anomaly score map to be optimized The designed foreground noise suppression module suppresses the influence of false anomalies in the foreground region during anomaly detection and obtains an anomaly score map with suppressed foreground noise. .

[0048] As shown in Figure 4, the foreground noise suppression module includes a Gaussian filter. This filter smooths the anomaly scores in local regions using a Gaussian kernel with a set size and variance, thereby reducing the anomaly scores of pseudo-anomaly regions on the foreground or object. The input to the foreground noise suppression module is the initial anomaly score map to be optimized. The output is an anomaly score map with foreground noise suppressed. The calculations for the entire process are as follows:

[0049]

[0050] in, This is a Gaussian filter function; in practice, its Gaussian kernel size can be set to [value missing]. The variance is 5.

[0051] S3, optimized for background anomaly suppression

[0052] For the industrial image to be detected as described in S1 above and the anomaly fraction map for suppressing foreground noise as described in S2 above... The resulting anomaly score map is further optimized through a background anomaly suppression module. Focusing more on anomalies in the foreground or object regions while suppressing the impact of pseudo-anomalies in the background regions, the optimized anomaly score is obtained. The background anomaly suppression module includes a non-background marked region extraction unit and a background anomaly removal unit;

[0053] As shown in Figure 5, the industrial image to be detected in S1 is input into the non-background marker region extraction unit, which outputs a marker map used to distinguish between the background and foreground regions. ;

[0054] The non-background labeled region extraction unit includes a pre-trained semantic segmentation model, Ground-SAM. Ground-SAM is based on the paper Ren T, Liu S, Zeng A, et al. Grounded sam: Assembling open-world models for diverse visual tasks[J]. arXiv preprint arXiv:2401.14159, 2024. The specific steps of the labeling process are as follows:

[0055] First, using pre-designed category-independent text cues, such as "foreground" and "object," the image to be detected is jointly input into the semantic segmentation model Ground-SAM to obtain multi-labeled maps related to the foreground cues. and multi-marker graphs related to object cues In a multi-label map, the value of relevant regions is 1, and the value of irrelevant regions is 0.

[0056] Then, the above multi-labeled images are merged by taking the union of the sets to obtain the initial non-background region labels, where the non-background region value is 1 and the background region value is 0.

[0057] Finally, the initial non-background regions are further labeled. The entire labeling process includes the following image operations: 1) 1) Remove small marked areas, which are usually marked areas that are not foreground or object areas; 2) 3) Remove small, noisy regions with holes, which are usually caused by errors in the semantic segmentation model itself; The dilation operation further ensures that all non-background regions are within the labeled area. The above image processing operations can be implemented using standard functions from the open-cv library in Python.

[0058] The final result is a marker map used to distinguish between background and foreground areas. The calculation process is as follows:

[0059]

[0060] As shown in Figure 6, the background anomaly removal unit uses the extracted marker map used to distinguish between the background and foreground regions. Further reduce the abnormal score graph Anomalies located in the background region. Specifically, a marker map used to distinguish between background and foreground regions. Anomaly score plot with suppressed foreground noise Element-wise multiplication is performed to remove outliers located in the background region of the outlier score map, resulting in an optimized outlier score map. Focusing on the foreground or object region where real anomalies exist, the calculation process is as follows:

[0061]

[0062] S4, Obtain the optimized anomaly score map

[0063] Using the anomaly score map obtained after step S3 Replace the original anomaly score map to be optimized This is used to improve the ability to identify / locate abnormal areas.

[0064] Figure 7 shows the effect of the proposed anomaly score map optimization method for industrial image anomaly detection, demonstrating the optimization results on normal samples, small anomaly region samples, and large anomaly region samples. Figure 7 shows the anomaly score map optimized by the method of this invention. Compared to the original anomaly score map to be optimized The anomaly scores of pseudo-anomaly regions in the foreground and background areas are effectively suppressed, while the impact on the anomaly scores of truly abnormal regions is relatively small, demonstrating that the anomaly score map optimization method for industrial image anomaly detection proposed in this invention has good performance.

[0065] Furthermore, this invention is not dependent on product category, requires no additional training, and is not limited to a specific anomaly detection method based on unsupervised or self-supervised learning. It can be used as a plug-and-play module.

[0066] In summary, this invention proposes an anomaly score map optimization method for industrial image anomaly detection. It primarily designs a foreground noise suppression module and a background anomaly suppression module. This invention takes the anomaly score map extracted by existing anomaly detection methods based on unsupervised or self-supervised learning as input. The foreground noise suppression module reduces the anomaly score of pseudo-anomalies in the foreground region, thereby suppressing the influence of pseudo-anomalies in the foreground region during anomaly detection. The background anomaly suppression module generates optimized labels related to the image foreground or object regions, ensuring that the optimized anomaly score map focuses only on the image foreground or object regions, thus suppressing the influence of pseudo-anomalies in the background region. Furthermore, this invention is independent of product category, requires no additional training, and is not limited to any specific anomaly detection method based on unsupervised or self-supervised learning; it can be used plug-and-play as a plugin. This invention further optimizes the anomaly score map extracted by anomaly detection methods based on unsupervised or self-supervised learning, thereby improving the ability to identify / locate anomaly regions.

[0067] Example 2: Based on the technical concept of the above method, the present invention also provides an anomaly score map optimization system for industrial image anomaly detection, including a foreground noise suppression module and a background anomaly suppression module. The foreground noise suppression module is used to suppress the influence of pseudo-anomalies in the foreground region during the anomaly detection process for the initial anomaly score map to be optimized, thereby obtaining an anomaly score map with suppressed foreground noise. The background anomaly suppression module is used to perform background anomaly suppression on the industrial image to be detected to obtain a marker map for distinguishing between the background and foreground regions. Based on the marker map for distinguishing between the background and foreground regions, outliers located in the background region in the anomaly score map are removed to obtain the optimized anomaly score map.

[0068] The foreground noise suppression module takes as input the initial anomaly score map to be optimized. The output is an anomaly score map with foreground noise suppressed. The calculation is as follows:

[0069]

[0070] in, Let be a Gaussian filter function, with a Gaussian kernel size of . The variance is 5.

[0071] The background anomaly suppression module includes a non-background marked region extraction unit and a background anomaly elimination unit;

[0072] The non-background labeled region extraction unit includes a pre-trained semantic segmentation model. The industrial image to be detected is used as input to this model. Pre-designed category-independent text prompts are combined with the industrial image to be detected and input into the trained semantic segmentation model to obtain multi-label maps related to the foreground prompts. and multi-marker graphs related to object cues In the multi-label map, the value of the relevant region is 1, and the value of the irrelevant region is 0. The above multi-label map is merged by taking the union to obtain the initial non-background region label, where the non-background region value is 1 and the background region value is 0.

[0073] The initial non-background region is processed as follows: mark regions without foreground or object elements are removed, hole noise regions are removed, and dilation is performed to obtain a marker map used to distinguish between background and foreground regions. ;

[0074]

[0075] The background anomaly removal unit uses the extracted marker map to distinguish between background and foreground regions. Anomaly score plot with suppressed foreground noise Perform element-wise multiplication to obtain the optimized anomaly score map. :

[0076] .

[0077] On the other hand, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the anomaly score map optimization method for industrial image anomaly detection described in the present invention.

[0078] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and the processor can implement the anomaly score map optimization method for industrial image anomaly detection described in the present invention when executing the computer executable program.

[0079] The computer device may be a laptop, a desktop computer, or a workstation.

[0080] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf programmable gate array (FPGA).

[0081] The memory described in this invention can be an internal storage unit of a laptop, desktop computer, or workstation, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0082] Computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0083] This invention has been described through embodiments. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, modifications can be made to these features and embodiments to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. An anomaly score map optimization method for industrial image anomaly detection, characterized in that, Includes the following steps: For the initial anomaly score map to be optimized, the influence of pseudo-anomalies in the foreground region during anomaly detection is suppressed to obtain an anomaly score map with suppressed foreground noise. Specifically, this includes: smoothing the anomaly scores in local regions of the initial anomaly score map to be optimized using a Gaussian kernel with set size and variance, reducing the anomaly scores in pseudo-anomaly regions on the foreground or objects, and obtaining an anomaly score map with suppressed foreground noise. Background anomaly suppression is performed on the industrial image to be detected to obtain a labeled map for distinguishing between background and foreground regions. Based on the labeled map for distinguishing background and foreground regions, outliers located in the background region in the anomaly score map are removed to obtain an optimized anomaly score map. Specifically, this includes: using pre-designed category-independent text prompts, and inputting the image to be detected into a trained semantic segmentation model to obtain multi-label maps related to the foreground prompts. and multi-marker graphs related to object cues The values ​​of relevant regions in the multi-label map are 1, and the values ​​of irrelevant regions are 0. The multi-label maps are merged by taking the union of the two maps to obtain the initial non-background region labels, where the non-background region values ​​are 1 and the background region values ​​are 0. The initial non-background region is processed as follows: the label regions without foreground or object are removed, the hole noise regions are removed, and the dilation operation is performed to obtain the label map used to distinguish the background and foreground regions. The label map used to distinguish the background and foreground regions is multiplied element-wise with the anomaly score map that suppresses foreground noise to obtain the optimized anomaly score map.

2. The anomaly score map optimization method for industrial image anomaly detection according to claim 1, characterized in that, Obtaining the initial anomaly score map to be optimized includes: For the industrial image data to be detected, an initial anomaly score map to be optimized is obtained by using existing anomaly detection methods based on unsupervised or self-supervised learning. The anomaly detection method based on unsupervised or self-supervised learning learns by modeling normal samples to obtain a model of the distribution or representation of normal images, and obtains an anomaly score map by calculating the difference between the sample to be detected and the modeled normal distribution or representation.

3. An anomaly fractional map optimization system for industrial image anomaly detection, characterized in that, It includes a foreground noise suppression module and a background anomaly suppression module. The foreground noise suppression module is used to suppress the influence of false anomalies in the foreground region during anomaly detection for the initial anomaly score map to be optimized, and obtain an anomaly score map with suppressed foreground noise. The background anomaly suppression module is used to suppress background anomalies in the industrial image to be detected, and obtain a marker map to distinguish between the background and foreground regions. Based on the marker map to distinguish between the background and foreground regions, outliers in the background region in the anomaly score map are removed to obtain an optimized anomaly score map. The foreground noise suppression module takes as input the initial anomaly score map to be optimized. The output is an anomaly score map with foreground noise suppressed. The calculation is as follows: in, It is a Gaussian filter function; The background anomaly suppression module includes a non-background marked region extraction unit and a background anomaly elimination unit; The non-background labeled region extraction unit includes a pre-trained semantic segmentation model. The industrial image to be detected is used as input to this model. Pre-designed category-independent text prompts are combined with the image to be detected and input into the trained semantic segmentation model to obtain multi-labeled maps related to the foreground prompts. and multi-marker graphs related to object cues ； In the multi-label map, the value of the relevant region is 1, and the value of the irrelevant region is 0; the above multi-label map is merged by taking the union to obtain the initial non-background region label, where the non-background region value is 1 and the background region value is 0; The initial non-background area is processed as follows: mark areas without foreground or object are removed, hole noise areas are removed, and a dilation operation is performed to obtain a marker map used to distinguish between background and foreground areas. ； The background anomaly removal unit uses the extracted marker map to distinguish between background and foreground regions. Marker map used to distinguish between background and foreground areas Anomaly score plot with suppressed foreground noise Perform element-wise multiplication to obtain the optimized anomaly score map. 。 4. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading part or all of the computer-executable program from the memory and executing it, and the processor executing part or all of the computer-executable program enabling the implementation of claim 1. The anomaly score map optimization method for industrial image anomaly detection as described in any one of the two claims.

5. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of claim 1. The anomaly score map optimization method for industrial image anomaly detection as described in any one of the two claims.

Images