Research On Patterned Fabric Defect Detection Algorithms Based On Feature Representation

Fabric defect detection is an important part of textile automation detection.The defect detection of plain or twill plain fabrics has been mature.The existing defect detection algorithms are not effective because of the variety of patterns,complex texture and different defect shapes.Aiming at the deficiency of defect detection of pattern fabric,this paper studies the defect detection algorithm of pattern fabric based on feature representation.The main contents and achievements are as follow:(1)The average hash feature descriptor is proposed to represent the average hash feature and gray level feature of pattern fabric,and the defect detection of pattern fabric is realized.Firstly,based on the periodicity of pattern fabrics,this paper constructs a average hash feature dictionary of defect-free fabrics.Then,the average hash feature and the gray feature in the test fabric image block are extracted;the average hash feature of the image block is matched with the dictionary,and the Hamming distance is used to represent the difference degree,and the structural saliency map is obtained.The grayscale feature of the image block is compared with the global grayscale average to obtain a grayscale saliency map.Finally,the saliency map is merged to locate the defect location.(2)A defect detection method based on non-negative matrix decomposition is proposed.Firstly,the period of the pattern is determined by the distance matching function,and the image block of the period size is taken as the detection sample.Then,the coefficient matrix of image block is obtained by non-negative matrix decomposition method and transformed into vector matrix,which is regarded as the unique feature representation of image block.The standard coefficient feature and similarity criterion are determined according to normal samples.Comparing the coefficients and standard coefficients of the test samples,the image block is judged to contain defects according to the established similarity criteria,and the fabric defect location is completed.In the process of non-negative matrix decomposition,non-negative double singular value decomposition initialization method is used to reduce the number of iterations of matrix decomposition.(3)A defect detection method based on convolutional self-encoder is proposed,and an unsupervised depth feature extraction model is established.Firstly,the periodic image block is selected as the input of the convolutional self-encoder model,and the Adam optimization algorithm,batch standardization and ReLU strategy are used to train the convolutional self-encoder model.Then the depth feature of the image block is extracted by the model.The defect location is determined by calculating the Euclidean distance between the features of normal fabric samples and those of fabric samples to be tested.In order to increase the correlation between image blocks represented by different locations,in the training process,image blocks of different locations are used as data input,and fixed target images are used as data output to realize multi-to-one mapping of image blocks.In view of the above three detection methods,this paper carries out experimental research,and through the test of fabric samples,verifies the effectiveness of the algorithm,which has a certain reference value for guiding practical application.