Research On Fabric Defect Detection System Using Machine Vision

Nowadays, the quality control of production plays an important role in the challenging commercial market. In textile industry, fabric defect is the primary factor affecting the quality of the fabrics and defect detection (also called inspection) is the main solution to fabric quality assurance. Traditional human inspection is labor-intensive and has low detection speed. Besides, the detection accuracy is affected by the experience and fatigue of the human inspector. Thus it is devoid of consistency and reliability. So an automatic inspection of fabric defect is a necessary and essential step of quality control in the textile manufacturing industry.In this dissertion, machine vision is introduced to textile industrial inspection field and used for visual inspection of fabric defects. Its key point is to use image processing and analysis methods to design a fast and accurate algorithm, which is capable of detecting all kinds of defects in different fabrics. Three metrics are used to measure the performance of the fabric detect detection algorithm including detection accuracy, universality (which means effective to all kinds of fabrics and defects), and real-time performance (which corresponds to the detection speed). And according to the objective of detection and expressional method of detection results, detect detection is divided into two types:detect discrimination and detect segmentation.In order to make the detection algorithm has a high degree of universality, which means effective to all kinds of fabric defects including global detects and local tiny ones, multi-resolution analysis method is used to analyse different fabric defects at different scales. Undecimated wavelet transform is used, instead of standard wavelet transform, for defect detection to achieve shift-invariance property. In order to increase the detection accuracy, a selection scheme of wavelet decomposition scales, which is capable of setting the decomposition scales adaptively to the spectral characteristics of fabric texture, is proposed to suppress the energy of normal texture and enhance the energy of detective regions. A simple and computationally effective data fusion scheme combined with amplitudes division of wavelet coefficients is used to fuse data from multiple scales together. And several features based on defective energy estimation are extracted from fused image, and these features are used for defect discrimination by thresholding.Because defect discrimination method can only locate defects in the fabric images, in order to acquire their morphological features such as size, length and orientation, the defect discrimination algorithm based on multi-resolution analysis is improved. A defect segmentation algorithm base on Gabor wavelet is proposed and an adaptive Gabor wavelet tuning method, which is capable of tuning the parameters of Gabor wavelet adaptively to the fabric texture, is also proposed. Compared to undecimated wavelet transform, Gabor wavelet transform can set the center of the pass-band of its filters more flexiblely. Thus the energy of normal texture is more efficiently suppressed and the energy of detective regions is more efficiently enhanced so that there is larger energy distinction between normal texture and defective regions. Thus after segmentation of filtered images by thresholding, higher segmentation accuracy can be acquired. Large quantities of fabric defect samples are implemented on the proposed defect segmatation algorithm as well as two other defect segmatation algorithms without adaptive tuning. Experimental results indicate that the proposed defect segmentation algorithm has better real-time performance and segmentation accuracy.To solve the problem of poor real-time performance in most fabric defect detection algorithms, a machine learning based defect discminination algorithm with high real-time performance is proposed. Two gray-level co-occurrence matrix features and two new textural features are extracted from the fabric image, and support vector data description classifier is used to classify the extracted features and dctcrmin whether they are defective or not. During the feature extraction progress, an adaptive quantization method is proposed to quantize256gray-levels into several quantized levels so that the computational complexity of feature extraction is trimendously reduced. Thus the real-time performance of the detection algorithm is grealty improved.The architecture of the fabric defect visual inspection system is elaborated. And the function and implementation methods of its components are described. Embedded digital signal processor is used instead of generic CPU to improve the computational power of the system. The image acquisition, transfer and processing method are also detailed. And a two-level Ping-Pong cache strategy is used to enable the image transfer and image processing running in parallel to impove the computational efficiency.