Research And Application Of Texture Image Anomaly Detection Based On Self-Supervised Learning

Many manufacturing companies in China have their quality inspection work done manually.Manual quality inspection is less efficient and prone to leakage.Machine vision-based quality inspection technology can well compensate for the lack of manual inspection.Machine vision-based quality inspection technology learns image features and performs quality inspection of product surfaces.The current mainstream machine vision-based quality inspection methods suffer from overfitting,excessive model size and lack of defect samples.These problems have caused great losses to both enterprises and society.In order to solve the above problems,this paper researches and analyzes the theory and methods for texture image anomaly detection.And the following work is carried out:(1)In order to solve the problem of difficult detection of minor anomalies,this paper proposes a data enhancement method based on CutPaste.The method uses CutPaste to construct negative samples with tiny anomalies.cutPaste randomly crops out thin rectangles or squares with small side lengths from normal samples and pastes the cropped rectangles randomly into any region of the original image.CutPaste also performs self-supervised learning as an auxiliary task.(2)In this paper,a texture image anomaly detection method(TIADSD)based on self-supervised learning and dictionary learning is proposed.The method mainly includes two stages:texture image feature representation and texture image anomaly detection.In the texture image feature representation stage.In order to improve the algorithm’s ability to detect minor anomalies,this paper proposes an image feature fitting method based on improved normalizing flow.The image features extracted by AlexNet are sampled and refitted using the normalizing flow,and the multiplicative coupling layer of the normalizing flow is increased to 10 layers.This improvement helps the normalizing flow to fit the features more accurately.In the texture image anomaly detection stage,in order that the overcomplete dictionary generated by dictionary learning can retain more image semantic information,this paper proposes an image anomaly detection method based on improved dictionary learning.By reducing the low-rankness constraint of dictionary learning,the representation is learned to an overcomplete dictionary containing only normal samples.The overcomplete dictionary is updated so that the overcomplete dictionary focuses more on the relevance of the image semantic information.To verify the performance of this method,the algorithm is fully validated and compared on texture data of MVTec AD in this paper.The experimental results show that the detection accuracy of the proposed texture image anomaly detection method in this paper is above 95%on different classes of texture data.In particular,the detection accuracy on the Carpet class reaches 99.7%,which is a 31%improvement compared to the baseline algorithm.In order to better verify the robustness of the texture image anomaly detection method,this paper also conducts experiments on the DAGM 2007 dataset.The experimental results show that the texture image anomaly detection method in this paper has good detection effect.The detection accuracy exceeds 80%on most of the texture data.The detection accuracies on Class2 and Class9 are over 95%.(3)In this paper,a texture image anomaly detection system is designed and implemented with the texture image anomaly detection method as the core.The system includes functions such as login interface,algorithm training and image detection.The algorithm embedded in the system can train the corresponding weights according to different categories of texture images,and use the corresponding weights to detect anomalies in texture images.The texture image anomaly method proposed in this paper has a high degree of innovation.The method can accurately detect the anomalous regions of texture images with strong robustness..