Abstract:

Aiming at the poor versatility of existing methods in various fabric defect types especially for warp and weft direction defects, this work presents unsupervised fabric defect segmentation using local texture feature. The proposed algorithm adopts unsupervised scheme, without need of any reference samples. For detection, the rarity of fabric defects is used to obtain local binary pattern (LBP) histogram features that can represent the local fabric texture from the entire image. Then, benefiting from the characteristics of woven fabrics' interlacing structure, and the one-dimension vectors obtained by projecting fabric image into warp and weft derections are extraced to represent local texture. Lastly, the anomaly maps of defect are computed from the extracted features, which are fused to sigment defect with weight factors used. The experimental results show that the proposed pwojection feature along warp and weft directions can well represent local fabric texture, which can achieve satiafied results in identifying defects by combining with LBP features.

Key words: fabric defect, local texture representation, warp and weft projection, anomaly detection