Unsupervised fabric defect segmentation using local patch approximation

In this work, a new method based on local patch approximation is presented to address automated defect segmentation on textile fabrics. The proposed method adopts unsupervised scheme without the need of reference images or any other prior information. Image patch is approximated by dictionary learned from a testing sample in the least squares sense. With the clue of the differentiation in approximation error, abnormal map (each pixel's anomalous likelihood) can be computed from the patch-level difference. The 2D maximum entropy with neighbourhood considered is applied to segment defective regions from the abnormal map. The experiments on 54 defective samples demonstrate that our method yields a robust and good overall performance with high precision and accepted recall rates.