Research On Fabric Defect Detection Algorithm Based On Visual Saliency

Fabric defect detection is the key link of fabric quality monitoring.Visual saliency can simulate human visual mechanism and quickly locate the targets with salient features.Therefore,textile defect detection based on visual saliency is of great research value.The research object of this paper is textiles with complex and periodic patterns.According to the periodic characteristics of textile images,the best block template is obtained by using the autocorrelation of images,which solves the problem of large periodic fluctuation of images extracted by traditional algorithms.Aiming at the deficiency that the original context visual saliency algorithm only considers the local saliency,it calculates the global and local saliency of adjacent pixel blocks at the same time.In order to improve the contrast between the defect and the background,the initial saliency map and a prior information are introduced into the low rank decomposition model.In order to deal with more scale information,double color space is used instead of single color space.This paper studies the key processes of fabric defect detection,such as image preprocessing and extracting visual salient features.The main research contents are as follows:（1）Research on textile defect detection method based on distortion correction and visual salient features.Firstly,according to the characteristics of textile image distortion,it can reduce the influence of textile image distortion and obtain the best block correction result according to the periodic distortion of textile image.Then the textile image is decomposed into texture layer and cartoon layer,and the improved context visual saliency algorithm is applied to obtain the saliency map of cartoon layer.Finally,the region growth method is used to segment the saliency map to separate the defect from the background,and then the morphological method is used to eliminate the small isolated points that are easy to cause false detection,so as to complete the defect detection process.Experiments show that the average recall rate of defect detection of three types of pattern textiles is 83.07%.（2）Research on textile defect detection method based on visual salient features in frequency domain.Aiming at the problem that the background texture of textile image interferes with the extraction of salient features,firstly,L₀ Gradient Minimization method is used to smooth the background texture of the image.Secondly,the textile image is represented in the form of quaternion image,and each pixel is represented by a quaternion composed of color,intensity and edge features.The saliency map of textile image is obtained by two-dimensional fractional Fourier transform.Finally,the region growth method is used for threshold segmentation to separate the defect and background,and then the morphological method is used to eliminate the small isolated points that are easy to cause false detection,so as to complete the defect detection process.Experiments show that compared with the six methods,this method has better average full and accurate results.（3）Research on textile defect detection method based on low-rank decomposition with multi-prior information and visual salient features.In order to improve the contrast between the defect and the background,the initial saliency map is constructed as the input of the low rank decomposition model.Then,local and global priors are fused to highlight the defect location.Then,the sparse part obtained by low rank decomposition is represented as a double four element image.Finally,quaternion Fourier transform and multi-scale Gaussian filter are used to generate multiple saliency maps to ensure the integrity of the detection results.Experiments show that the average recall rate of this method for the defect detection of three types of pattern textiles is 84.89%.