Research On Unsupervised Fabric Defect Detection Algorithm Based On Image Retrieval

Printed fabrics not only contain complex texture features but also have rich pattern information.It has much better visual effects than solid fabrics and has become one of the most popular textile products.However,the failure of printing machine,yarn/textile defection,and many other factors will cause various defects with significant differences both in shape and physical sizes in the surface of the printed fabric.Traditional manual inspection methods or automatic defect detection method for solid fabrics cannot meet printed fabric production requirements due to low accuracy and poor stability.And automatic detection of printing defect technology is significant for improving printing fabrics' appearance and quality.In recent years,it has attracted wide attention from industry and academia.Most of the current fabric defect detection algorithms rely on the analysis of defect-free samples,and they are mainly facing the following challenges: 1.accurately extracting the periodical primitive.2.Pixel level printed fabrics image reconstruction.3.The highly dependence of deep learning method on training set.In order to solve the above problems,we build a content based image retrieval(CBIR)system,and trained Convolutional Denoising Auto-Encoder(CDAE)and Hash Encoder(HE)in unsupervised manner for feature extraction and indexing.Intention of finding a defect-free reference image with the same resolution,consistent and aligned pattern as the test image by the way of image retrieval.And defect detection stage is based on this reference image.Our proposed image retrieval system includes a database of printed fabric local patterns,a CDAE and a HE.The database of printed fabric local patterns includes image database,feature vector set,hash code table.The feature vector set is obtained by input the elements of image database into CDAE,and then input all of feature vectors into HE to get the hash code table.The CDAE is a convolutional neural network the total number of parameters is only 8 million together with the HE,which is far more less than the method using object detection and GAN networks,and there is also no need to collect defective samples for training,which is especially beneficial when it is difficult to collect defective samples.To make the database of printed fabric local patterns containing all features of the test printed fabric patterns,we proposed a selection criteria of sample image,and then a sliding box is applied to extract patches with equal size from the sample image to build the image database.The CDAE prevents the system affecting by the texture of the fabric and provides a reliable feature description of the patterns.The HE indexes the feature vectors to binary code while maintaining their similarityWith the retrieved reference image,the defect is determined by applying the Tsallis entropy thresholding and opening operation on the difference map.Experimental results demonstrate the effectiveness and the efficiency of the proposed method in the defect detection for printed fabric with complex pattern.