Fabric defect detection via low-rank decomposition with gradient information and structured graph algorithm

• Published in: Information sciences Vol. 546; pp. 608 - 626
• Main Authors: Shi, Boshan, Liang, Jiuzhen, Di, Lan, Chen, Chen, Hou, Zhenjie
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 06.02.2021
• Subjects: Cycle information; Defect detection; Graph algorithm; Low-rank decomposition; Prior knowledge; Prior knowledge; Low-rank decomposition; Cycle information; Graph algorithm; Defect detection
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2020.08.100

The low rank decomposition model shows the potential of fabric defect detection, in which a matrix is decomposed into a sparse matrix representing the defect free region (background) and identifying the defect area (foreground). However, there are still two shortcomings. Firstly, when the texture of defect image has high gradient feature, the sparse matrix obtained by the existing model still retains a large number of edges of undetected regions. Secondly, due to the imprecision of prior information, most models will misjudge the defect free points around the defect block when dealing with the small defect area or the defect area containing multiple cycles. In order to solve these problems, this paper proposes a fabric defect detection method based on low rank decomposition of gradient information and structured graph algorithm: 1) structured graphics algorithm, according to the characteristics of fabric defect image, fabric defect image is divided into defect free block with local feature and defect damage period. 2) In the merging process, an adaptive threshold is set according to the number of cycles contained in the current block to encourage intra lattice merging and prevent the merging of defective blocks and surrounding non defective blocks. 3) The defect prior information calculated from the segmentation results is used to guide matrix decomposition to weaken the defect free region and highlight the defect area under the sparse term. We evaluated our model on a standard database and compared it with the four latest methods. The total TPR and fpr of this method are 87.3% and 1.21% respectively on box, star and point databases, which is the best performance.