Unsupervised fabric defects detection based on spatial domain saliency and features clustering

Fabric defects detection plays a critical role in the quality control of textile manufacturing industry. It is still a challenge to realize accurate fabric defects detection due to variations of fabric texture and the lack of defective samples. To solve this problem, this paper proposes an unsupervised learning fabric defects detection method. Firstly, a multi-level spatial domain saliency method (MSDS) is proposed to generate multi-level saliency values by convoluting color histograms with pixel values, which can greatly suppress background information via the fusion of multi-level saliency values. Secondly, fabric feature extraction method (FFE) is proposed to respectively extract geometrical features, intensity features, and texture features from potential defective regions. Finally, an adaptive fabric feature clustering algorithm (AFFC) is designed to adjust weights of fabric features and obtain final defects detection results. In the experiment section, the influence of fabric features on defects detection is discussed. And compared with other unsupervised learning methods, the proposed method can achieve over 90% accuracy fabric defects detection within small samples, which is significantly better than other methods and can meet the practical requirements of fabric defects detection.