Defect Detection Algorithm of Periodic Texture by Multi-metric-Multi-module Image Voting Method

Abstract

Appling deep learning network models to train and detect defects on periodic texture background images requires a large number of standard datasets. However, in the field of texture fabric defect detection, there is lack of public standard datasets, and it is pretty time-consuming and laborious to prepare a high-quality training dataset. In this study, we propose a comprehensive method combining with the characteristics of periodic texture images, which uses multiple metrics and multiple mathematical models of an image to vote and score the splitted sub-images of the image, so as to detect the locations of defects on the periodic texture image. Central to our method is subimage segmentation, Zero-Slope-RANSac(ZS-RANSac) method, Multi-metric-Multi-model Image Voting strategy, which utilizes the local consistency of image metrics existing in periodic texture by cutting images into sub-images of the same size. To obtain the basic scoring matrix of each sub-image under each model, we take the difference of the standard value of the non-defect background calculated by ZS-RANSac and all measurements of sub-image, and then combine the matrix and multiple numeration model. According to the order of the scores, a certain proportion of polymer image points are considered as outer points, which are the defect sub-images. This method completely relies on statistical strategy to make use of the periodic texture characteristics of the image, and can detect the non-lattice texture image without training data. It has a wide application prospect for the textile industry, which requires real time and lacks high-quality training datasets.