A visual attention system based anisotropic diffusion method for an effective textile defect detection

Abstract

Recently, saliency maps are widely used as an image feature to improve results of segmentation or classification of defects. In this paper we tackle the problem of defect detection in fabric images. We propose a defect extraction method in textile images based on a modified anisotropic diffusion filter combined with a saliency image feature. The classical anisotropic diffusion models consider only the image gradient information of a diffused pixel. Consequently, they cannot identify defect edges confused with background texture. Since the defects in the neighborhood of the image generally have larger saliency intensity value than the background, the proposed diffusion model incorporates both local gradient magnitude and a modified saliency map. The aim is to preserve defect edges while effectively removing the textured background. Moreover a new diffusion function is proposed that permits to carry on adaptively a smoothing in faultless regions and a sharpening process at defect boundaries. To confirm the effectiveness of our algorithm, we compare tests and results with different methods on textile images including different kinds of defects and fabrics. Results show that our method outperforms the other considered algorithms and can accurately extracts defects.