Research Of Fabric Defect Detection Based On Hybrid Self-adaptive Wavelet Basis

Fabric defect detection is an import step to guarantee the textile quality. In the process of textile fabric production, the appearance of defects in fabric is inevitable and it can reduce the value of textile greatly. As a result, it may lead to the waste of human and material resources. Therefore, to detect the defects in fabric effectively and make accurate classification of textiles according to the fabric quality have important application value and significance to the textile industry.However, the fabric texture is complicated and various, and its defect shape and categories are random. Therefore, how to detect accurately the size and location of the fabric defect from the fabric image becomes the research focus and the research content in this field. Traditional fabric defect methods are Based on the detection method of feature extraction and non-feature extraction. They still contain defect information in the reconstruction of fabric image texture and the defect detection rate is low. To improve the detection precision of fabric defect, this thesis is based on sparse method of the fabric defect detection algorithms. First, we learn the adaptive dictionary library from fabric image to be detected and use the dictionary to represent the original fabric image sparsely; Then, the sparse coefficient matrix is calculated, and the reconstruction normal fabric textile can be achieved by using dictionary and the coefficient matrix; and then, the residual image can be generated by subtracting the reconstruction image and original fabric image; finally, the residual image is segmented to locate the defects area. This thesis respectively proposed the defect detection methods based on sparse optimization, based on sparse representation of the MLBP features and based on HOG feature sparse subspace classification, and the research performances are as follows:(1) For sparse representation based defect detection method, the reconstructed normal fabric image contains some defects, and it will result in reducing detection accuracy. In this thesis, a novel fabric defect detection algorithm based on sparse optimization is proposed. Firstly, an adaptive dictionary is learned from test fabric image using L1-norm minimization method, the test fabric image is sparsely represented using the learned dictionary, and then we calculate the coefficient matrix of sparse representation; secondly, we remove the abnormal coefficients using optimization function, then a new image is reconstructed using the optimized coefficient matrix and the dictionary; finally, the reconstructed image is subtracted from original test image to get a residual image, then the maximum entropy threshold method is used to segment the defect region. Experimental results demonstrate the proposed algorithm has higher detection accuracy comparing with the state of the art.(2) Fabric defect detection is essential in textile quality control. The thesis presents a fabric defect detection algorithm based on main local binary pattern(MLBP) extraction and sparse representation method. The proposed algorithm extracts the main local binary pattern of an image and establishes a dictionary base. The dictionary base atom is used to sparsely represent the LBP of the central pixel of a non-overlapping 3 × 3 image block and calculate the mean value of the gray-scale difference between the corresponding pixels of the main local binary pattern(MLBP) image and the neighborhood pixels in the 3 × 3 image block. The algorithm gives consideration to the mean value of the gray-scale difference, the sparse representation coefficient matrix, and the gray level of the central pixel, and reconstructs the original fabric image. The algorithm finally calculates the residual between the reconstructed image and the original fabric image that contains defect, obtains the residual image, adopts the maximum entropy threshold segmentation method to divide the residual image, and finds the defective region. The experiment results indicate that the proposed algorithm that combines MLBP operator and sparse representation can reconstruct the fabric image of the normal texture and the defective region derived from the calculation can effectively highlight the residual image and achieve better defect detection capability.(3) This thesis proposed a fabric defects detection algorithm based on HOG features extraction and sparse representation. First, the 256×256 pixels fabric images is divided into blocks with the same size of 16×16; then, the HOG features are extracted from each block, and the dictionary is built with these features; the 1-norm minimization method is adopted to realize the sparse representation of HOG feature of each block, and the sparse representation coefficient matrix can be calculated; finally, the sparse subspace clustering is used to classify all the blocks as normal block or defect block according to the their coefficient difference. The division result and the defects location can be achieved. The experience performance shows that the proposed method can better build the dictionary and reconstruct the original fabric texture, and the results is more exact than other methods with the highlighted defects area.