Research On Fabric Defect Detection Algorithm Based On Hierarchical Feature And Improved RPCA

Fabric defect detection plays an essential role in the process of textile manufacturing,which directly determines the quality and value of textiles.For the fabric images with complex and diverse textures and different morphologies of defects,the traditional pattern recognition method has a poor adaptability and low detection accuracy.Robust principal component analysis(RPCA),also known as low-rank decomposition model,can divide the image into target and background,and can be applied to fabric defect detection.However,the performance of the detection method based on low-rank decomposition model depends on the effective representation of the image and the construction and solution of the model.Therefore,this thesis mainly researches image representation and RPCA model combining with the peculiarity of fabric images,and presents the fabric defect detection algorithm based on the hierarchical features and the improved RPCA.The research results are as follows:1)The fabric defect detection algorithm based on feature fusion and TV-RPCA is proposed.Firstly,Canonical Correlation Analysis(CCA)was adopted to integrate the1st-order gradient feature and 2st-order gradient feature to improve the image representation ability.Then,a low-rank decomposition model based on the total variation regularization term is constructed,which can not only divide the fabric defects effectively,but also eliminate part of the noise in the fabric image to a certain extent.Finally,the sparse matrix obtained by optimization was used to obtain the defect saliency map according to the spatial correspondence relation,and then threshold segmentation operation was performed.Experiment results show that the proposed algorithm is superior to detection effect based on single feature and traditional RPCA.2)The fabric defect detection algorithm based on deep feature and NTV-RPCA is proposed.Firstly,a convolutional neural network is adopted to extract multi-leveldeep features to solve the representation problem of fabric images.Then,low-rank decomposition model based on non-convex total variation regularization term is constructed,which can not only effectively detect the defect saliency map with less noise,but also improve the solution accuracy due to non-convex optimization.Finally,a selective fusion is performed on the multiple defect saliency maps generated by the sparse matrix,and the final defect segmentation map is obtained by threshold segmentation operation.Experiment results show that the proposed algorithm further improves the defects detection effect.3)The fabric defect detection algorithm based on deep-low-level feature and NTV-NRPCA is proposed.Firstly,the high-level semantic information extracted by a new convolutional neural network and some low-level contrast information are fused to improve the image representation ability.Then,non-convex low-rank decomposition model based on non-convex total variation regularization term is constructed,which can not only effectively detect the defect saliency map with less noise,but also further improve the solution accuracy.Finally,the defect saliency map obtained by the sparse matrix is sent into a threshold segmentation operation to locate the defect location.Experiment results show that the proposed algorithm still has high detection accuracy under the premise of reducing computational complexity.The research results of this thesis can be applied to the detection of fabric image defects with simple or complex textures,and improves the adaptability and detection accuracy of existing detection methods.The algorithm can be popularized to detect defects on the surface of paper,aluminum foil or steel and other industrial products,and has an extensive application prospect.