Detection Of Diverse Fabric Defects Via Modified Robust PCA

Patterned fabric is manufactured by a set of predefined symmetry rules.In the production process and entering the market must go through a variety of inspection and test,the fabric defect detection is the most important part which aim is to detect whether the fabric by the flaws and defects in position.At the moment,there are some automatic detection machine,the task is mainly done by manual offline detection,while this method is low test efficiency and test results are greatly influenced by cloth inspection personnel’s subjective,the inefficient detection rate and the higher residual error rate,Since the above reasons,the fabric defects automatic detection is a common concern for scholars at home and abroad in recent years.The main task of fabric defect detection is to estimate the position of a defective area in a fabric map and output a gray image and a binary image.The pixel value of each point in the gray scale indicates the possibility that the pixel belongs to the flaw;A binary graph indicates whether each pixel is a defect.Considering that a patterned fabric is manufacture used by a set of symmetry rules,it can be assumed as the superposition of defective regions(spares structure)and repeated patterns(principal component low-rank structure).Robust Principal Component Analysis(Robust PCA)can be used to recover low-rank matrix so that it can be used to detect various fabric images in general.Whereas,fabric images are generally contaminated by some noise caused by the fabrics’ distortion and illumination changes in the process of image acquisition.Inaccurate results are likely to be generated for the existence of these noise.To handle the problem,we improve the Robust PCA model with a noise term measured by matrix F-norm to suppress the noise.Since Robust PCA and N-RPCA can’t handle some low-rank defective regions in the theory.We propose a defective prior,thus utilize texture feature to extract the preliminary defect for N patches which is introduced to guide the decomposition process.Comparison on detection results and efficiency about diverse patterned fabric images with various defects demonstrate that our method is more efficient and more robust as for the parameters and noise.