Unsupervised defect detection method and device based on deep learning

Abstract

The invention provides an unsupervised defect detection method and equipment based on deep learning, and belongs to the technical field of defect detection.The method comprises the steps that defect detection is conducted on an image to be detected, and a coarse defect image containing a defect area is obtained; calculating a first similarity between the coarse defect image and each defect image in a pre-constructed defect image set; calculating a second similarity between the coarse defect image and each false detection defect image in a pre-constructed false detection defect image set; and performing defect authenticity detection on the coarse defect image according to the first similarity and the second similarity. By accurately calculating the similarity, real defects and false detection conditions can be effectively distinguished, and the detection precision can be remarkably improved; moreover, compared with a large amount of labeled data required by a deep learning model, the method focuses on direct comparison of existing samples, so that the dependence on a large-scale labeled data set is reduced, and the defect identification efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of defect detection technology, and in particular to an unsupervised defect detection method and device based on deep learning. Background Art

[0002] Wafer defect detection is an essential process in the field of semiconductor production. The wafer defect detection process requires the ability to output the defect location based on each frame of the image to be inspected. The difficulty lies in the fact that due to the instability of optical conditions during shooting, there will be problems such as distortion, bright and dark fields. At the same time, due to the differences in manufacturing, some models will have more pseudo defects, which cannot be well filtered from grayscale and saturation. Therefore, the traditional Golden Die (the golden template used for comparison in the traditional grain detection process, referring to the appearance of the grain in a defect-free state) detection method will face many limitations.

[0003] In order to solve the problem of complex defect characteristics and get rid of the dependence on Golden Die accuracy, existing solutions use deep learning methods to detect defects. However, due to the limitations of the working environment of the semiconductor industry, there is no clear definition of various defects and corresponding defect sample libraries. Most solutions use semi-supervised and unsupervised network structures, but semi-supervised and unsupervised defect recognition methods still face considerable challenges in terms of defect recognition accuracy and efficiency. Summary of the invention

[0004] The present invention provides an unsupervised defect detection method and device based on deep learning, which are used to solve the defects of low accuracy and efficiency of defect recognition in the prior art.

[0005] In a first aspect, the present invention provides an unsupervised defect detection method based on deep learning, comprising: performing defect detection on an image to be inspected, and obtaining a coarse defect image containing a defect area; calculating a first similarity between the coarse defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the coarse defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed-detected defect images, which are missed-detected defect images, and falsely detected defect images are pseudo-defect images that identify normal areas as defect areas; and performing defect authenticity detection on the coarse defect image based on the first similarity and the second similarity.

[0006] According to the unsupervised defect detection method based on deep learning provided by the present invention, defects are detected on the image to be inspected to obtain a rough defect image containing a defect area, including: using a deep learning model to extract a feature map of the image to be inspected; calculating the cosine similarity between each pixel in the feature map and all other pixels, so as to take the maximum cosine similarity as the self-similarity coefficient of each pixel; taking pixels whose self-similarity coefficient is lower than a preset similarity threshold as abnormal points; constructing a rough defect area according to the abnormal points in the feature map, and obtaining a rough defect image.

[0007] According to the unsupervised defect detection method based on deep learning provided by the present invention, a coarse defect area is constructed according to the abnormal points in the feature map, and a coarse defect image is obtained, including: determining the abnormal area in the feature map according to the abnormal points in the feature map; when the number of pixels in the abnormal area is greater than a preset pixel number threshold, treating the abnormal area as the coarse defect area; and generating the coarse defect image according to the coarse defect area.

[0008] According to the unsupervised defect detection method based on deep learning provided by the present invention, the authenticity of the defect is detected on the rough defect image according to the first similarity and the second similarity, including: when the first similarity is greater than a first preset threshold, determining that the defect identified in the image to be inspected is a real defect; when the second similarity is greater than a second preset threshold, determining that the defect identified in the image to be inspected is a false defect.

[0009] According to the unsupervised defect detection method based on deep learning provided by the present invention, after performing defect detection on the image to be inspected and obtaining a rough defect image containing the defect area, it also includes: when it is determined that a missed detection has occurred based on the rough defect image, marking the missed detection area on the image to be inspected through human-computer interaction; generating a missed detection defect image based on the missed detection area, and updating the defect image set; and, when it is determined that a false detection has occurred based on the rough defect image, marking the false detection area on the image to be inspected through human-computer interaction; generating a pseudo defect image based on the false detection area, and updating the false detection defect image set.

[0010] According to the unsupervised defect detection method based on deep learning provided by the present invention, the deep learning model is a Visual Transformers model.

[0011] The unsupervised defect detection method based on deep learning provided by the present invention also includes: adjusting the first preset threshold and the second preset threshold with the goal of the defect detection accuracy being greater than the preset accuracy.

[0012] In a second aspect, the present invention also provides an unsupervised defect detection device based on deep learning, comprising: performing defect detection on an image to be inspected, and obtaining a rough defect image containing a defect area; calculating a first similarity between the rough defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the rough defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed-detected defect images, which are missed-detected defect images, and falsely detected defect images are pseudo-defect images that identify normal areas as defect areas; and performing defect authenticity detection on the rough defect image based on the first similarity and the second similarity.

[0013] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-described unsupervised defect detection methods based on deep learning are implemented.

[0014] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described unsupervised defect detection methods based on deep learning.

[0015] In a fifth aspect, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned unsupervised defect detection methods based on deep learning.

[0016] The unsupervised defect detection method and device based on deep learning provided by the present invention can effectively distinguish between real defects and false detections by accurately calculating the similarity. Especially when dealing with a large number of pseudo-defects, this method can significantly improve the detection accuracy. Moreover, compared with the large amount of labeled data required by the deep learning model, this method focuses more on the direct comparison of existing samples, which reduces the dependence on large-scale labeled data sets and improves the defect recognition efficiency.

[0017] The present invention provides a deep learning-based defect detection method that combines multiple steps such as feature extraction, cosine similarity calculation, and outlier recognition, which can effectively improve the accuracy and robustness of defect detection and is very suitable for high-precision detection fields such as semiconductor wafers.

[0018] The present invention adopts the backbone of the segmentation task as the feature extractor and re-performs semantic segmentation training on the ADE dataset, so that the speed of feature extraction is greatly improved. In addition, the separability of the segmentation backbone for local features under the cosine metric is better, and the controllability in parameter tuning is enhanced.

[0019] The present invention can improve the accuracy of defect detection by continuously iteratively adjusting the first preset threshold and the second preset threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is a flowchart of the unsupervised defect detection method based on deep learning provided by the present invention;

[0022] Figure 2 It is a schematic diagram of a process of obtaining a rough defect image containing a defect area provided by the present invention;

[0023] Figure 3 It is a schematic diagram of defect misdetection provided by the present invention;

[0024] Figure 4 It is a schematic diagram of misdetected area marking through human-computer interaction provided by the present invention;

[0025] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] It should be noted that, in the description of the embodiments of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0028] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more.

[0029] Combine the following Figure 1-Figure 5 The present invention describes an unsupervised defect detection method and device based on deep learning provided by an embodiment of the present invention.

[0030] Figure 1 is a flow chart of the unsupervised defect detection method based on deep learning provided by the present invention, such as Figure 1 As shown, including but not limited to the following steps:

[0031] Step 101: perform defect detection on the image to be inspected, and obtain a rough defect image including the defect area.

[0032] The image to be detected in the present invention may be a wafer image, but is not limited to a wafer image; the rough defect image is a "rough" image containing possible defect areas, which may contain real defects or falsely detected defects. The present invention may use image processing techniques such as edge detection, threshold segmentation, feature extraction, etc. to identify potential defect areas.

[0033] Step 102: Calculate a first similarity between the coarse defect image and each defect image in a pre-constructed defect image set; and calculate a second similarity between the coarse defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set.

[0034] The defect image set includes missed defect images and other real defect images. The missed defect images are defect images that are missed, and the falsely detected defect images are false defect images that identify normal areas as defect areas.

[0035] The present invention quantitatively measures the similarity between the coarse defect image and each image in the defect image set by calculating the first similarity; similarly, the present invention quantitatively measures the similarity between the coarse defect image and each image in the falsely detected defect image set by calculating the second similarity.

[0036] Algorithms for calculating similarity include, but are not limited to, structural similarity index (SSIM), normalized cross correlation (NCC), histogram comparison, and cosine similarity, which will not be described in detail here.

[0037] Step 103: performing defect authenticity detection on the rough defect image according to the first similarity and the second similarity.

[0038] Whether a defect is real or fake is determined based on the comparison result of the first similarity and the second similarity. For example, if a rough defect image has a high similarity with an image in the defect image set, but has a low similarity with all images in the false detection defect image set, the defect can be considered to be real.

[0039] The unsupervised defect detection method based on deep learning provided by the present invention can effectively distinguish between real defects and false detections by accurately calculating the similarity. Especially when dealing with a large number of pseudo-defects, this method can significantly improve the detection accuracy. Moreover, compared with the large amount of labeled data required by the deep learning model, this method focuses more on the direct comparison of existing samples, which reduces the dependence on large-scale labeled data sets and improves the defect recognition efficiency.

[0040] Figure 2 FIG. 1 is a flow chart of obtaining a rough defect image including a defect area provided by the present invention, such as Figure 2 As shown, including but not limited to the following steps:

[0041] Step 201: Extract a feature map of the image to be inspected using a deep learning model.

[0042] Use a pre-trained deep learning model (such as a convolutional neural network (CNN)) to extract feature representations of the image to be inspected.

[0043] The feature map reflects the important features of different areas in the image and provides a basis for subsequent defect detection.

[0044] Step 202: Calculate the cosine similarity between each pixel in the feature map and all other pixels, and use the maximum cosine similarity as the self-similarity coefficient of each pixel.

[0045] Cosine similarity is used to measure the angular similarity between two vectors. The value range is usually between -1 and 1, where 1 means that the two vectors are in the same direction, -1 means they are in opposite directions, and 0 means they are orthogonal.

[0046] For each pixel, the maximum cosine similarity between it and all other points is selected as the self-similarity coefficient of the point. This step helps to identify relatively unique or abnormal areas in the feature map.

[0047] Step 203: Pixels whose self-similarity coefficients are lower than a preset similarity threshold are regarded as outliers.

[0048] A preset similarity threshold is set, and pixels with a self-similarity coefficient lower than the threshold are marked as outliers.

[0049] Step 204: construct a coarse defect area according to the abnormal points in the feature map, and obtain a coarse defect image.

[0050] The identified abnormal points are connected to form connected areas, which are coarse defect areas. These areas are cut out from the original image to obtain a coarse defect image.

[0051] The present invention provides a deep learning-based defect detection method that combines multiple steps such as feature extraction, cosine similarity calculation, and outlier recognition, which can effectively improve the accuracy and robustness of defect detection and is very suitable for high-precision detection fields such as semiconductor wafers.

[0052] Specifically, the present invention can better capture complex patterns in images by extracting features through a deep learning model, thereby improving the robustness of detection; using cosine similarity as the basis for identifying outliers, it can effectively distinguish normal areas from abnormal areas and improve the accuracy of defect detection; the method can adapt to different defect types and sizes, and can flexibly respond to different application scenarios by adjusting parameters such as the similarity threshold; by setting a reasonable similarity threshold, the possibility of mistakenly marking normal areas as defects can be reduced.

[0053] Based on the contents of the above embodiments, as an optional embodiment, the unsupervised defect detection method based on deep learning provided by the present invention constructs a coarse defect area according to the abnormal points in the feature map and obtains a coarse defect image, including the following steps.

[0054] Step 301: Determine an abnormal area in the feature map according to abnormal points in the feature map.

[0055] The present invention can use a connected component analysis algorithm to find out the area formed by adjacent abnormal points.

[0056] Step 302: When the number of pixels in the abnormal region is greater than a preset pixel number threshold, the abnormal region is regarded as the rough defect region.

[0057] Set a preset pixel number threshold, such as 50 pixels.

[0058] For each abnormal region, if the number of its pixels exceeds 50, it is marked as a coarse defect region.

[0059] Step 303: Generate the coarse defect image according to the coarse defect area.

[0060] For each region marked as a coarse defect area, mark or crop it on the original image. If it is marking, a bounding box can be drawn around each region; if it is cropping, the part containing these regions is directly cropped from the original image to obtain a coarse defect image.

[0061] Based on the content of the above embodiment, as an optional embodiment, the unsupervised defect detection method based on deep learning provided by the present invention performs defect authenticity detection on the rough defect image according to the first similarity and the second similarity, including: when the first similarity is greater than a first preset threshold, determining that the defect identified in the image to be inspected is a real defect; when the second similarity is greater than a second preset threshold, determining that the defect identified in the image to be inspected is a false defect. The following is a specific description of this process:

[0062] When the first similarity is greater than the first preset threshold, this indicates that the defect in the image to be inspected is very similar to a known real defect, and therefore the identified defect can be considered to be real.

[0063] When the second similarity is greater than the second preset threshold, this indicates that the defect in the image to be inspected is very similar to a known pseudo defect, and therefore the identified defect can be considered to be a pseudo defect.

[0064] Assume that the first preset threshold is 0.7 and the second preset threshold is 0.6. If the first similarity is greater than 0.7, the identified defect is considered to be a real defect. If the second similarity is greater than 0.6, the identified defect is considered to be a pseudo defect.

[0065] Based on the contents of the above embodiments, as an optional embodiment, the unsupervised defect detection method based on deep learning provided by the present invention, after performing defect detection on the image to be inspected and obtaining a rough defect image containing the defect area, also includes: in a case where it is determined that a missed detection has occurred according to the rough defect image, marking the missed detection area on the image to be inspected through human-computer interaction; generating a missed detection defect image based on the missed detection area, and updating the defect image set; and, in a case where it is determined that a false detection has occurred according to the rough defect image, marking the false detection area on the image to be inspected through human-computer interaction; generating a pseudo defect image based on the false detection area, and updating the false detection defect image set.

[0066] The present invention can allow the user to view the detection results by developing a human-computer interaction interface. The user can mark the missed detection or false detection area on this interface, generate new defect images or pseudo-defect images according to the user's marking, and update the data set.

[0067] For example, Figure 3 Schematic diagram of defect misdetection provided by the present invention, such as Figure 3 As shown, Figure 3 The sharp corners framed in the upper middle area are the normal area, i.e., the false detection area; Figure 4 is a schematic diagram of misdetected area marking through human-computer interaction provided by the present invention, such as Figure 4 As shown, Figure 4The misdetected areas in the image are marked to facilitate the subsequent generation of pseudo defect images.

[0068] As an optional embodiment, the present invention provides an unsupervised defect detection method based on deep learning, wherein the deep learning model is a Visual Transformers model.

[0069] The Visual Transformers (ViT) model is a visual model based on the Transformer architecture, which divides the image into a series of fixed-size patches and treats these patches as sequences input into the Transformer for processing.

[0070] Different from other schemes that use the classification backbone as the feature extractor, the present invention uses the backbone of the segmentation task as the feature extractor and re-performs semantic segmentation training on the ADE dataset, which greatly improves the speed of feature extraction. In addition, the segmentation backbone has better separability for local features under the cosine metric and has stronger controllability in parameter tuning.

[0071] Based on the content of the above embodiment, as an optional embodiment, the unsupervised defect detection method based on deep learning provided by the present invention further includes: adjusting the first preset threshold and the second preset threshold with the goal of the accuracy of defect detection being greater than the preset accuracy. The following is an optional step of the present invention.

[0072] Step 1: Initialize the threshold

[0073] An initial value, such as 0.5, is selected for the first preset threshold and the second preset threshold.

[0074] Step 2: Calculate similarity

[0075] The features of the rough defect image, defect image and pseudo-defect image of the image to be inspected are extracted using the ViT model.

[0076] A first similarity between the rough defect image feature vector and the real defect image feature vector is calculated. A second similarity between the rough defect image feature vector and the pseudo defect image feature vector is calculated.

[0077] Step 3: Evaluate performance

[0078] The authenticity of the defect is determined based on the first similarity and the second similarity in combination with the first preset threshold and the second preset threshold, and the accuracy of defect detection is calculated.

[0079] Step 4: Threshold Adjustment

[0080] The threshold is adjusted according to whether the accuracy of defect detection is greater than the preset accuracy, and steps 2 and 3 are repeated until the target accuracy is reached.

[0081] If the accuracy of defect detection is less than the preset accuracy, the threshold needs to be adjusted. The present invention can try to improve the accuracy by increasing or decreasing the threshold.

[0082] Step 5: Output the results

[0083] When the preset accuracy is reached, the optimal first preset threshold and second preset threshold are output. The present invention can improve the accuracy of defect detection by continuously iteratively adjusting the first preset threshold and the second preset threshold.

[0084] Figure 5 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520 and the memory 530 communicate with each other through the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute the unsupervised defect detection method based on deep learning, the method comprising: performing defect detection on the image to be inspected, obtaining a coarse defect image containing a defect area; calculating a first similarity between the coarse defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the coarse defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed defect images, the missed defect images are missed defect images, and the falsely detected defect images are falsely detected defect images that identify normal areas as defect areas; and performing true and false defect detection on the coarse defect image according to the first similarity and the second similarity.

[0085] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0086] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the unsupervised defect detection method based on deep learning provided by the above-mentioned embodiments, the method including: performing defect detection on an image to be inspected to obtain a rough defect image containing a defect area; calculating a first similarity between the rough defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the rough defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed defect images, the missed defect images are missed defect images, and the falsely detected defect images are pseudo defect images that identify normal areas as defect areas; based on the first similarity and the second similarity, performing defect authenticity detection on the rough defect image.

[0087] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the unsupervised defect detection method based on deep learning provided in the above-mentioned embodiments, the method comprising: performing defect detection on an image to be inspected, and obtaining a coarse defect image containing a defect area; calculating a first similarity between the coarse defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the coarse defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed-detected defect images, which are missed-detected defect images, and falsely detected defect images are pseudo-defect images that identify normal areas as defect areas; and performing defect authenticity detection on the coarse defect image based on the first similarity and the second similarity.

[0088] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An unsupervised defect detection method based on deep learning, characterized in that: include: Detect defects on the image to be inspected, and obtain a rough defect image containing the defect area; Calculating a first similarity between the rough defect image and each defect image in a pre-constructed defect image set; and calculating a second similarity between the rough defect image and each falsely detected defect image in a pre-constructed falsely detected defect image set; the defect image set includes missed-detected defect images, which are missed-detected defect images, and falsely detected defect images are pseudo-defect images that identify normal areas as defect areas; The authenticity of defects in the rough defect image is detected according to the first similarity and the second similarity.

2. The unsupervised defect detection method based on deep learning according to claim 1, characterized in that: Detect defects on the image to be inspected and obtain a rough defect image containing the defect area, including: Use deep learning models to extract feature maps of images to be inspected; Calculate the cosine similarity between each pixel in the feature map and all other pixels, and use the maximum cosine similarity as the self-similarity coefficient of each pixel; Pixels whose self-similarity coefficient is lower than a preset similarity threshold are regarded as abnormal points; A coarse defect area is constructed according to the abnormal points in the feature map, and a coarse defect image is obtained.

3. The unsupervised defect detection method based on deep learning according to claim 2, characterized in that: Constructing a coarse defect area according to the abnormal points in the feature map and obtaining a coarse defect image, including: Determine an abnormal area in the feature map according to the abnormal points in the feature map; When the number of pixels in the abnormal area is greater than a preset pixel number threshold, the abnormal area is regarded as the rough defect area; The rough defect image is generated according to the rough defect area.

4. The unsupervised defect detection method based on deep learning according to claim 1, characterized in that: According to the first similarity and the second similarity, detecting the authenticity of defects on the rough defect image includes: When the first similarity is greater than a first preset threshold, determining that the defect identified in the image to be inspected is a real defect; When the second similarity is greater than the second preset threshold, it is determined that the defect identified in the image to be inspected is a pseudo defect.

5. The unsupervised defect detection method based on deep learning according to claim 1, characterized in that: After performing defect detection on the image to be inspected and obtaining a rough defect image containing the defect area, the following steps are also included: In the case where missed detection is determined according to the rough defect image, marking the missed detection area on the image to be inspected in the form of human-computer interaction; generating a missed detection defect image according to the missed detection area, and updating the defect image set; and When it is determined that a false detection occurs according to the rough defect image, a false detection area is marked on the image to be detected through human-computer interaction; a pseudo defect image is generated according to the false detection area, and the false detection defect image set is updated.

6. The unsupervised defect detection method based on deep learning according to claim 2, characterized in that: The deep learning model is a Visual Transformers model.

7. The unsupervised defect detection method based on deep learning according to claim 4, characterized in that: Also includes: With the goal of making the accuracy of defect detection greater than the preset accuracy, the first preset threshold and the second preset threshold are adjusted.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the unsupervised defect detection method based on deep learning as described in any one of claims 1 to 7 are implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the unsupervised defect detection method based on deep learning as described in any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the unsupervised defect detection method based on deep learning as described in any one of claims 1 to 7 are implemented.

Images