The Fabric Defect Detection Algorithm Based On Linear CCD

Currently, with the rapid development of textile industry, the quality of the cloth in themarket increases higher and higher. Defect detection has thus become one of the importantlinks in the textile quality control. In the traditional cloth inspection process, the defectdetection is mostly done by the workers with the naked eye, this approach is not only theexistence many shortcomings of low efficiency, accuracy, high rate of undetected, but alsoaffected by many factors.With the computer science and digital imaging technology, the issue based on thedomestic fabric defect detection proposes and verifies a detection algorithm based on linearCCD. The algorithm is divided into three parts: One is the linear CCD timing controlalgorithm; the other is fabrics feature extraction algorithm; the third is the defect categorydetection algorithm. These three parts can all be applied to the fabric defect detection system,allowing the image acquisition and the image processing modules of the detection system tobetter co-ordination. The algorithm can short the processing time of the system and increasethe classification accuracy.The main subject of work done: First, based on the internal structure and timingcharacteristics of TCD1206SUP, using the hardware description language of VHDL andFPGA top-down modular design, complete the driving pulse of the linear CCD design; andthen, doing the wavelet decomposition and reconstruction after the image pre-processing ofthe fabric in order to extract the statistic characteristics of the radial and the latitudinal detailsof the image; finally, an improved defect classification algorithm is proposed highlighted. Thealgorithm is named by the classification algorithm of BP neural network based on CART tree.The method compared to the traditional BP neural network fabric defect classificationalgorithm, firstly, using the CART tree to do feature selection, extraction of the high degree ofdistinction between the characteristics to dimension reduction of the feature parameters, andtraining the BP neural network with the feature after dimensionality reduction to achieve thepurpose of defect classification. The CART tree is the most important step in thisclassification algorithm comparing of the traditional BP neural network. After validation, themethod not only reduces the overall network training cycles, but also greatly improve theclassification speed and accuracy.