Research And Development Of Real-time Vision-based Fabric Defect Detection System

Our country is a competitive textile powerful nation. However in the market system, the textile production is still in a low stage. With the increased cost of employment, manufacture and production demand, and the quality standards improved, we can not only rely on human to guarantee the quality control of the production efficiency and the product. And because of the subjective factors, the inspector may make a low rate of efficiency and high rate of false and missing detection, which unable to adapt to the rapid development of industrial demand. So it’s urgent to develop the fabric detection system which can practically meet the actual needs of the industrial field.In this paper, the grey fabric is detected as the main target. On the basis of detailed analysis of the existing technical, an experimental platform is built with an industrial line scan camera a swept source and other hardware, and with the improved algorithm based on machine vision and image processing, which is suitable for the production line of fabric defects detection. According to the texture characteristics of the fabric image, this paper mainly research on segmentation of defects, machine learning classification algorithms, and make the improvement and optimization.Firstly, this paper discuss the image filtering, sharpening and gray-scale range adjustment during the image acquisition process, which belongs to the image preprocessing of automatic detection for fabric defects. And make a comparison with several threshold segmentation algorithms, to choose the appropriate one to separate the defect.Secondly,this paper analyzes the response characteristics of the Gabor filter for texture image, and then puts forward an improved model based on wavelet neural networks, which chooses Gabor filter to replace the traditional incentive function, by the Levenberg-Marquardt optimization algorithm to realize the optimization. The even and the odd parts of the Gabor filter are used respectively to detect defects and then integrated. Meanwhile, this paper calculates the gray level co-occurrence matrix and other image feature information in the defects region, and classifies the fabric image by support vector machine according to the 4 point system.Finally, the paper builds a cartoon-texture decomposition on fabric texture image, and makes improvement on the total variation minimization model which could have a response on defects part, and at the same time eliminate the interference of background texture. The experiment indicates that this method can effectively divide the texture and the defects.