Unsupervised texture class defect detection method and system based on salient object detection

Abstract

The application provides a kind of unsupervised texture class defect detection method and system based on salient target detection, comprising the following steps: S1, combine the images in the texture class data set to generate new defect images; S2, build a salient target detection network framework and extract image features to generate a feature map of texture data; S3, obtain the feature map of normal images and the normal image feature vector at each pixel in the target texture class data, forming a normal image feature dictionary; S4, train the network according to the target texture class data containing defects and its label, and obtain a pre-trained model of the salient target detection network; S5, obtain a pseudo-normal image feature map and use the normal image feature map for supervision to generate a final defect detection result map. The method provided by the application only needs to train the model with part of the label of the known texture class data set, and can realize unsupervised defect detection of the target texture class industrial product, reducing the cost and improving the work efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect detection, in particular to a texture defect detection method and system based on salient object detection. BACKGROUND

[0002] Industrial defect detection is a technology applied in industrial production, aiming to detect and identify defects and abnormalities in the manufacturing process. Traditional industrial defect detection is mainly divided into three categories: texture feature-based method, color feature-based method, and shape feature-based method. Texture features reflect the homogenization phenomenon in the image, which can be reflected by the gray distribution of pixels and their nearby spatial neighborhood. Among them, the color feature-based method has small calculation amount, small dependence on the size, direction and angle of the image, and high robustness; the shape-based method effectively utilizes the target of interest in the image for retrieval; and the texture feature-based method obtains the shape parameters of the image by describing the outer boundary features of the object, which is the main method type.

[0003] At present, the requirement for product quality in industrial production is getting higher and higher. However, in the process of product manufacturing, some defective products are often produced, such as steel plates, aluminum foils, and ceramic tiles, which are usually texture-based industrial products and often have defects such as scratches, spots, and holes. These defects not only affect the appearance of the product, but also have adverse effects on the use of the product. Therefore, defect detection is crucial in the production of texture-based industrial products.

[0004] The rapid development of deep learning has made it more and more widely used in the field of defect detection.

[0005] There are three common classifications based on deep learning: supervised method, unsupervised method, and weakly supervised method.

[0006] Among them, the supervised method requires both training set and test set, and the samples in the training set must be labeled; the training set is used to find the internal law of the samples, and then the law is applied to the test set. The unsupervised method learns the patterns of unlabeled data to obtain some internal features and relationships of the data and automatically classifies the data. The weakly supervised method combines the characteristics of supervised and unsupervised methods, and can balance the labeling cost and performance according to specific needs.

[0007] Salient object detection aims to detect the most visually salient objects in a scene and extract their precise pixel-level contours. The salient object detection model is not sensitive to the category, and the salient objects detected by it are universal and irrelevant to the category. Texture-based industrial products have the characteristics of simple background and repeated structure.

[0008] However, traditional salient object detection relies on large-scale data training and a large number of datasets to train models, and its performance in terms of boundary blurring, practicality, and accuracy is unsatisfactory.

[0009] Chinese patent document CN107622280B discloses a method for detecting surface defects in leather materials based on feature modeling saliency detection. The method includes the following steps: S01, firstly, acquiring video information of the leather material and selecting several frames from the video as images to be detected; S02, then using a bottom-up model to initially acquire saliency images, obtaining a first saliency image; S03, enhancing the first saliency image using an equal-area multi-histogram equalization algorithm, obtaining a second saliency image; S04, processing the second saliency image based on superpixel re-optimization, obtaining a third saliency image; S05, finally, marking the location of defects based on the third saliency image, and adaptively thresholding the obtained third saliency image using a global thresholding method. While this patent document can detect and locate defects in the entire piece of leather at once, it still requires a large-scale training dataset, thus failing to solve the aforementioned problem. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide an unsupervised method and system for detecting texture defects based on salient target detection.

[0011] An unsupervised texture-based defect detection method based on salient target detection according to the present invention includes:

[0012] Step S1: Extract the defective portion from the defective image in the texture dataset as the foreground image, use the normal image of the corresponding category as the background image, and combine the foreground image and the background image to generate a new defective image;

[0013] Step S2: Based on the known texture data, build a salient object detection network framework and extract image features to generate a feature map of the texture data;

[0014] Step S3: Obtain the feature map of the normal image in the target texture class data; obtain the normal image feature vector at each pixel to form a normal image feature dictionary;

[0015] Step S4: Train the network based on the defective target texture class data and its labels to obtain the pre-trained model of the salient target detection network;

[0016] Step S5: Based on the detection results obtained from the pre-trained model, replace the features of the defective regions in the image with the feature vectors in the normal image feature dictionary to obtain a pseudo-normal image feature map. Use the normal image feature map for supervision to generate the final defect detection result map.

[0017] Preferably, step S1 includes the following sub-steps:

[0018] Step S1.1: Divide the defective and normal images in the texture dataset into training and test sets; use the labeled known texture data as the dataset used for pre-training the model;

[0019] Step S1.2: Divide the labels in the training set into edge labels and center region labels, and use the unlabeled texture class dataset as the target texture class defect dataset;

[0020] Step S1.3: Extract the defect portion from the defect images in the training set as the foreground image;

[0021] Step S1.4: Collect normal images of the corresponding category of the foreground image and use them as background images;

[0022] Step S1.5: Combine the foreground image and the background image to generate a new defect image and expand the defect dataset.

[0023] Preferably, step S2 includes the following sub-steps:

[0024] Step S2.1: Input the image of size (b,c,w,h) into the salient object detection network framework; where b represents the number of samples processed at the same time, c represents the number of feature channels of each sample, h represents the height of the image, and w represents the width of the image;

[0025] Step S2.2: Use the ResNet-50 network as the backbone network to extract image features; the ResNet-50 network outputs multi-layer features F, F = {F...} i |i=1,2,3,4,5····};The feature size of the i-th layer is Among them, c i This represents the number of channels in the i-th layer feature; the multi-layer feature F obtains image edge feature maps and image center region feature maps through convolutional layers, and the feature size of the i-th layer is... Where c is the pre-defined number of channels, and i = 2, 3, 4, 5, ...;

[0026] Step S2.3: Integrate the edge feature map and the central region feature map of the image, and generate a complete feature map after linearization.

[0027] Preferably, the pre-trained model of the salient object detection network is obtained by training the network using defect images and their labels from known texture class data, and the salient object detection network framework is obtained by training the network using known texture class data.

[0028] Preferably, step S5 includes the following sub-steps:

[0029] Step S5.1: Use a pre-trained model to process the target texture class defect data samples and obtain a rough defect detection result image;

[0030] Step S5.2: Replace the features identified as defective regions in the results with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map;

[0031] Step S5.3: Use normal image feature maps to supervise pseudo-normal image feature maps, adjust network parameters to generate the final defect detection result map.

[0032] An unsupervised texture-based defect detection system based on salient target detection according to the present invention includes:

[0033] Module M1: Extracts the defective portion from the defective image in the texture dataset as the foreground image, uses the normal image of the corresponding category as the background image, and combines the foreground image and the background image to generate a new defective image;

[0034] Module M2: Based on known texture data, build a salient object detection network framework and extract image features to generate feature maps of the texture data;

[0035] Module M3: Obtains the feature map of the normal image in the target texture class data; obtains the normal image feature vector at each pixel to form a normal image feature dictionary;

[0036] Module M4: Trains a network based on the defective target texture data and its labels to obtain a pre-trained model of the salient target detection network;

[0037] Module M5: Based on the detection results obtained from the pre-trained model, the features of the defective regions in the image are replaced with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map. The normal image feature map is then used for supervision to generate the final defect detection result map.

[0038] Preferably, module M1 includes the following sub-modules:

[0039] Module M1.1: Divides defective and normal images in the texture dataset into training and test sets; uses labeled known texture data as the dataset for pre-training the model;

[0040] Module M1.2: Divides the labels in the training set into edge labels and center region labels, and uses the unlabeled texture class dataset as the target texture class defect dataset;

[0041] Module M1.3: Extracts the defect portion from the defect images in the training set as the foreground image;

[0042] Module M1.4: Collects normal images of the corresponding category of the foreground image and uses them as background images;

[0043] Module M1.5: Combines foreground and background images to generate new defect images, expanding the defect dataset.

[0044] Preferably, module M2 includes the following sub-modules:

[0045] Module M2.1: Inputs an image of size (b,c,w,h) into the salient object detection network framework; where b represents the number of samples processed simultaneously, c represents the number of feature channels per sample, h represents the height of the image, and w represents the width of the image;

[0046] Module M2.2: Uses a ResNet-50 network as the backbone network to extract image features; the ResNet-50 network outputs multi-layer features F, F = {F...} i |i=1,2,3,4,5····};The feature size of the i-th layer is Among them, c i This represents the number of channels in the i-th layer feature; the multi-layer feature F obtains image edge feature maps and image center region feature maps through convolutional layers, and the feature size of the i-th layer is... Where c is the pre-defined number of channels, and i = 2, 3, 4, 5, ...;

[0047] Module M2.3: Integrates the edge feature map and the central region feature map of the image, and generates a complete feature map after linearization.

[0048] Preferably, the pre-trained model of the salient object detection network is obtained by training the network using defect images and their labels from known texture class data, and the salient object detection network framework is obtained by training the network using known texture class data.

[0049] Preferably, module M5 includes the following sub-modules:

[0050] Module M5.1: Uses a pre-trained model to process target texture-type defect data samples and obtains a coarse defect detection result image;

[0051] Module M5.2: Replaces the features identified as defective regions in the results with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map;

[0052] Module M5.3: Uses normal image feature maps to supervise pseudo-normal image feature maps, and adjusts network parameters to generate the final defect detection result map.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The method and system provided by this invention do not rely on large-scale data training. They only require the labels of a portion of known texture datasets to train the model, which can be transferred to the defect detection of other types of target texture industrial products. This enables unsupervised defect detection of target texture industrial products, reduces costs, improves work efficiency, and has strong generalization ability.

[0055] 2. The method and system provided by this invention improve the problem of blurred boundaries in the prior art, and can realize more refined texture defect detection, which can be widely applied in the field of industrial surface defect detection with diverse scenarios.

[0056] 3. The method and system provided by this invention incorporate normal image features of target texture class on the basis of the pre-trained model for salient target detection, which can be easily deployed on industrial production lines and simultaneously has real-time performance, accuracy and robustness.

[0057] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0058] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0059] Figure 1 This is a flowchart of an unsupervised texture-based defect detection method based on salient target detection in an embodiment of the present invention. Detailed Implementation

[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0061] Reference Figure 1 As shown, an unsupervised texture-based defect detection method based on salient object detection includes:

[0062] First, the defective and normal images in the texture class dataset are divided into training and test sets, and the labeled known texture class data is used as the dataset for pre-training the model. The labels in the training set are divided into edge labels and center region labels, and the unlabeled texture class dataset is used as the target texture class defect dataset.

[0063] Defective parts are cropped from defective images in the training set as foreground images. Normal images of the corresponding categories of the foreground images are collected and used as background images. Then, the foreground images and background images are combined to generate new defective images, thus expanding the defective dataset.

[0064] Next, an image of size (b,c,w,h) is input into the salient object detection network framework; where b represents the number of samples processed at the same time, c represents the number of feature channels for each sample, h represents the height of the image, and w represents the width of the image.

[0065] The ResNet-50 network is used as the backbone network to extract image features. The ResNet-50 network outputs multi-layer features F, where F = {F...} i |i=1,2,3,4,5····};The feature size of the i-th layer is Among them, c i This represents the number of channels in the i-th layer feature; then, the multi-layer feature F obtains the image edge feature map and the image center region feature map through convolutional layers, and the feature size of the i-th layer is... Where c is the pre-defined number of channels, and i = 2, 3, 4, 5, ...;

[0066] Next, the edge feature map and the central region feature map of the image are integrated and linearized to generate a complete feature map.

[0067] First, obtain the feature maps of normal images from the target texture class data, then obtain the normal image feature vector at each pixel to form a normal image feature dictionary. Train the network based on the defective target texture class data and its labels to obtain a pre-trained model for the salient object detection network.

[0068] Finally, a pre-trained model is used to process the target texture defect data samples to obtain a coarse defect detection result map. The features of the defective regions in the result are replaced with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map. The pseudo-normal image feature map is supervised by the normal image feature map, and the network parameters are adjusted to generate the final defect detection result map.

[0069] The above are basic embodiments of the present invention. The technical solution of the present invention will be further described below through a preferred embodiment.

[0070] Reference Figure 1 As shown, the dataset is first preprocessed and augmented.

[0071] The specific preprocessing operations for the dataset are as follows: the labeled texture data of leather, mesh, carpet, and tile in the MVTec anomaly detection dataset are used as known texture defect datasets and used as the training set for the first-stage network; the labels in the training set are divided into edge labels and central region labels, and the wood board dataset is used as the target texture defect dataset.

[0072] The specific operations for data augmentation are as follows: Defective portions are extracted from the defect images in the defect datasets for leather, mesh, carpet, and tile categories and used as the foreground image for synthesizing new defect images. Corresponding normal images from the leather, mesh, carpet, and tile categories are selected and used as the background image for synthesizing new defect images. Then, a portion of the defects from the foreground image is randomly combined with the background image to generate new defect images, thus completing the augmentation of the defect dataset.

[0073] This step uses the training and test sets of five texture classes from the MVTec anomaly detection dataset, including leather, mesh, carpet, tile, and wood.

[0074] First, pre-trained models using data from leather, mesh, carpet, and tile categories are used to perform unsupervised defect detection on wood board data. Then, defective parts are extracted from defective images of leather, mesh, carpet, and tile categories, and normal images of the corresponding categories are collected as foreground and background images, respectively. The foreground and background images are then overlaid to augment the datasets of leather, mesh, carpet, and tile categories.

[0075] Next, we will build a salient object detection network.

[0076] The image is divided into edge and center regions. A ResNet-50 network is used for feature extraction. The features from the 2nd, 3rd, 4th and 5th layers of the extracted features are processed by convolutional layers to obtain the features of the edge and center regions, respectively. Two decoders are set up to decode these two types of features to obtain the edge feature map and the center region feature map. Then the edge feature map and the center region feature map are concatenated to generate the feature map of the whole image.

[0077] Specifically, an image with dimensions (b, c, w, h) is first input into the salient object detection network. Here, b represents the batch size, i.e., the number of samples processed simultaneously, which is set to 32; c represents the number of channels, i.e., the number of feature channels for each sample; and w and h represent the width and height of the image, i.e., the image size of each sample, respectively.

[0078] Next, a ResNet-50 network is used as the backbone network for feature extraction. The ResNet-50 network can output five layers of features, denoted as F = {F...}i |i=1,2,3,4,5}, the feature size of the i-th layer is Among them, c i This represents the number of channels in the i-th layer feature.

[0079] In subsequent steps, features from layers 2, 3, 4, and 5 of the five-layer feature set obtained from ResNet are used. These four layers are then processed through two convolutional layers to extract two sets of features for the edge and center regions. The feature size of the i-th layer is... Where c is the pre-set number of channels, and i = 2, 3, 4, 5.

[0080] Then, the same operation is performed on the features of the edge and center regions, namely, linear interpolation upsampling of the features of layers 3, 4, and 5 and fusing them with the features of layer 2 to generate two features of size 1. Features and connections to form a size of The feature map is linearized to obtain the final saliency map.

[0081] The feature maps of normal images in the target texture class are obtained through the backbone network ResNet-50, and the feature vector of normal image at each pixel is obtained to form a normal image feature dictionary.

[0082] Taking wooden boards as an example:

[0083] First, features of normal images in the plank class are extracted using the ResNet-50 backbone network, and the feature maps of layers 2, 3, 4, and 5 are stored. Then, the normal image feature vectors at each pixel in the feature maps of layers 2, 3, 4, and 5 are stored to form four normal image feature dictionaries.

[0084] This step only requires a small number of normal, unlabeled images from the target texture class data to achieve unsupervised defect detection of defective images of the target texture class.

[0085] Finally, the network in the salient object detection network framework is trained using normal images, defective images, and their labels from the leather, mesh, carpet, and tile data in the MVTec anomaly detection dataset, resulting in a pre-trained model.

[0086] Let's take wood planks as an example:

[0087] During the training process for the target texture data, a feedback mechanism is established:

[0088] First, a pre-trained model is used to process data samples of wooden board defects to obtain a coarse defect detection result image. Then, the features extracted by the backbone network are replaced by randomly replacing the features of the defective regions identified in the coarse defect detection result image with feature vectors from the normal image feature dictionary, generating a pseudo-normal image feature map. Finally, the normal image feature map is used to supervise the pseudo-normal image feature map and adjust the network parameters to generate the final defect detection result image.

[0089] The present invention also provides an unsupervised texture defect detection system based on salient target detection. The unsupervised texture defect detection system based on salient target detection can be implemented by executing the process steps of the unsupervised texture defect detection method based on salient target detection. That is, those skilled in the art can understand the unsupervised texture defect detection method based on salient target detection as a preferred embodiment of the unsupervised texture defect detection system based on salient target detection.

[0090] Specifically, an unsupervised texture-based defect detection system based on salient object detection includes:

[0091] Module M1: Extracts the defective portion from the defective image in the texture dataset as the foreground image, uses the normal image of the corresponding category as the background image, and combines the foreground image and the background image to generate a new defective image;

[0092] Module M2: Based on known texture data, build a salient object detection network framework and extract image features to generate feature maps of the texture data;

[0093] Module M3: Obtains the feature map of the normal image in the target texture class data; obtains the normal image feature vector at each pixel to form a normal image feature dictionary;

[0094] Module M4: Trains a network based on the defective target texture data and its labels to obtain a pre-trained model of the salient target detection network;

[0095] Module M5: Based on the detection results obtained from the pre-trained model, the features of the defective regions in the image are replaced with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map. The normal image feature map is then used for supervision to generate the final defect detection result map.

[0096] The module M1 includes the following sub-modules:

[0097] Module M1.1: Divides defective and normal images in the texture dataset into training and test sets; uses labeled known texture data as the dataset for pre-training the model;

[0098] Module M1.2: Divides the labels in the training set into edge labels and center region labels, and uses the unlabeled texture class dataset as the target texture class defect dataset;

[0099] Module M1.3: Extracts the defect portion from the defect images in the training set as the foreground image;

[0100] Module M1.4: Collects normal images of the corresponding category of the foreground image and uses them as background images;

[0101] Module M1.5: Combines foreground and background images to generate new defect images, expanding the defect dataset.

[0102] Module M2 includes the following sub-modules:

[0103] Module M2.1: Inputs an image of size (b,c,w,h) into the salient object detection network framework; where b represents the number of samples processed simultaneously, c represents the number of feature channels per sample, h represents the height of the image, and w represents the width of the image;

[0104] Module M2.2: Uses a ResNet-50 network as the backbone network to extract image features; the ResNet-50 network outputs multi-layer features F, F = {F...} i |i=1,2,3,4,5····};The feature size of the i-th layer is Among them, c i This represents the number of channels in the i-th layer feature; the multi-layer feature F obtains image edge feature maps and image center region feature maps through convolutional layers, and the feature size of the i-th layer is... Where c is the pre-defined number of channels, and i = 2, 3, 4, 5, ...;

[0105] Module M2.3: Integrates the edge feature map and the central region feature map of the image, and generates a complete feature map after linearization.

[0106] The pre-trained model of the salient object detection network is obtained by training the network using defect images and their labels from known texture data, and the salient object detection network framework is obtained by training the network using known texture data.

[0107] The module M5 includes the following sub-modules:

[0108] Module M5.1: Uses a pre-trained model to process target texture-type defect data samples and obtains a coarse defect detection result image;

[0109] Module M5.2: Replaces the features identified as defective regions in the results with feature vectors from the normal image feature dictionary to obtain a pseudo-normal image feature map;

[0110] Module M5.3: Uses normal image feature maps to supervise pseudo-normal image feature maps, and adjusts network parameters to generate the final defect detection result map.

[0111] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An unsupervised texture class defect detection method based on salient object detection, characterized in that, Comprising: Step S1: intercepting the defect part in the defect image in the texture class data set as a foreground image, taking the normal image corresponding to the category as a background image, and combining the foreground image and the background image to generate a new defect image; Step S2: according to the known texture data, a salient object detection network framework is built, a first stage pre-training model is obtained, image features are extracted, and a feature map of the texture data is generated; Step S3: obtaining the feature map of the normal image in the target texture class data; obtaining the normal image feature vector at each pixel to form a normal image feature dictionary; Step S4: training the network according to the target texture class data containing defects and its label, and obtaining a second stage pre-training model of the salient object detection network; Step S5: using the pre-training model to process the target texture class defect data sample, obtaining a rough defect detection result image, replacing the features of the defect area in the result image with the feature vectors in the normal image feature dictionary, obtaining a pseudo normal image feature map, and using the normal image feature map for supervision to generate the final defect detection result image. 2.The unsupervised texture class defect detection method based on salient object detection of claim 1, wherein, The step S1 comprises the following sub-steps: Step S1.1: dividing the defect image and the normal image of the texture class data set into a training set and a test set; taking the known texture class data with labels as a data set used when pre-training the model; Step S1.2: intercepting the defect part in the defect image in the training set as a foreground image; Step S1.3: collecting the normal image of the corresponding category of the foreground image as a background image; Step S1.4: combining the foreground image and the background image to generate a new defect image and expand the defect data set.

3. The unsupervised texture class defect detection method based on salient object detection according to claim 1, wherein, The step S2 comprises the following sub-steps: Step S2.1: inputting an image with a size of into a salient object detection network framework; wherein, represents the number of samples processed simultaneously, represents the number of feature channels of each sample, represents the height of the image, represents the width of the image; Step S2.2: extracting image features by taking a ResNet-50 network as a backbone network; the ResNet-50 network outputs multi-layer features F, ; the first layer has a feature size of , wherein represents a channel number of the first layer feature; the multi-layer features F obtain an image edge feature map and an image center region feature map through a convolution layer, and the first layer has a feature size of ; wherein ; Step S2.3: integrating the edge feature map and the center region feature map of the image, and generating a complete feature map after linearization processing.

4. The unsupervised texture class defect detection method based on salient object detection according to claim 1, wherein, The first stage pre-training model of the salient object detection network is obtained by training the network using the defect image and its label of the known texture class data, and the salient object detection network framework is trained using the known texture class data.

5. The unsupervised texture class defect detection method based on salient object detection according to claim 1, wherein, The step S5 comprises the following sub-steps: Step S5.1: using the pre-training model to process the target texture class defect data sample, obtaining a rough defect detection result image; Step S5.2: replacing the features of the defect area in the result image with the feature vectors in the normal image feature dictionary to obtain a pseudo normal image feature map; Step S5.3: using the normal image feature map to supervise the pseudo normal image feature map, adjusting the network parameters to generate the final defect detection result image.

6. An unsupervised texture class defect detection system based on salient object detection, characterized in that, Comprising: Module M1: intercepting the defect part in the defect image in the texture class data set as a foreground image, taking the normal image corresponding to the category as a background image, and combining the foreground image and the background image to generate a new defect image; Module M2: according to the known texture data, a salient object detection network framework is built, a first stage pre-training model is obtained, image features are extracted, and a feature map of the texture data is generated; Module M3: obtaining the feature map of the normal image in the target texture class data; obtaining the normal image feature vector at each pixel to form a normal image feature dictionary; Module M4: training the network according to the target texture class data with defects and its label, and obtaining a pre-trained model of the second stage of the salient object detection network; Module M5: processing the target texture class defect data sample using the pre-trained model, obtaining a rough defect detection result image, replacing the features of the defect area in the result image with the feature vectors in the normal image feature dictionary to obtain a pseudo-normal image feature map, and using the normal image feature map for supervision to generate a final defect detection result image.

7. The unsupervised texture class defect detection system based on salient object detection according to claim 6, wherein, The module M1 includes the following sub-steps: Module M1.1: dividing the defect images and normal images of the texture class data set into a training set and a test set; and using the known texture class data with labels as a data set for use in the pre-trained model; Module M1.2: cutting the defect part from the defect image in the training set as a foreground image; Module M1.3: collecting normal images corresponding to the category of the foreground image as background images; Module M1.4: combining the foreground image and the background image to generate a new defect image, and expanding the defect data set.

8. The unsupervised texture class defect detection system based on salient object detection according to claim 6, wherein, The module M2 includes the following sub-steps: Module M2.1 : inputting an image with a size of into a salient object detection network framework; wherein, represents the number of samples processed simultaneously, represents the number of feature channels of each sample, represents the height of the image, represents the width of the image; The module M2.2 extracts image features by taking a ResNet-50 network as a backbone network; the ResNet-50 network outputs multi-layer features F, ; the first layer has a feature size of , wherein, ; the first layer has a feature size of ; wherein ; and ; Module M2.3: integrating the edge feature map and the center region feature map of the image, and generating a complete feature map after linearization processing.

9. The unsupervised texture class defect detection system based on salient object detection according to claim 6, wherein, The pre-trained model of the first stage of the salient object detection network is obtained by training the network using the defect images and their labels of the known texture class data, and the salient object detection network framework is trained using the known texture class data.

10. The unsupervised texture class defect detection system based on salient object detection according to claim 6, wherein, The module M5 includes the following sub-steps: Module M5.1: processing the target texture class defect data sample using the pre-trained model to obtain a rough defect detection result image; Module M5.2: replacing the features of the defect area in the result image with the feature vectors in the normal image feature dictionary to obtain a pseudo-normal image feature map; Module M5.3: using the normal image feature map to supervise the pseudo-normal image feature map, adjusting the network parameters to generate a final defect detection result image.

Images