A Study of Automatic Inspection of Fabric Defects

• Main Authors: Su-Wei Huang, 黃淑慧
• Other Authors: Tung-Hsu（Tony） Hou
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/36825318116030118411

碩士 === 國立雲林科技大學 === 工業工程與管理研究所 === 87 === Manufacturing processes of the textile industry are already highly automated, but the inspection process is still performed by operators using human visual inspection. Therefore, some computer vision based systems were developed in these years to replace human operators and to automate the inspection process. These systems, by using Fourier transforms, gray-level statistics and morphology analysis, co-occurrence matrix analysis, were trying to detect and classify fabric defects and texture appearance. However, these systems were suffered from their low speed and low recognition rate. In this research, a new system will be developed. The system is composed of the following procedures: image acquisition, image binarization, noise reduction, image projection, feature extraction, and the back propagation neural network. The major feature of the system is to combine image projection with the back propogation neural network to detect and classify fabric defects. The image projection procedure takes the projection histograms of a binarized image along the X and Y directions. The coefficient of kurtosis, will then be extracted and fed to the back propagation neural network. Due to its simple processing, speed of the proposed system is expected to be very fast. In order to demonstrate the system's performance, the proposed system will be compared with the typical systems which used the co-occurrence matrix analysis and the back propagation neural network, and the fast Fourier transformation and the back propagation neural network. The defects to be inspected are broken threads in the longitude, broken threads in the latitude, holes in the fabrics, oil stains in the fabrics, and hooky fabrics. This research finds that the proposed system has 100% inspection accuracy, while the co-occurrence matrix method has 76.6% accuracy and the fast Fourier transformation method has 91.4% accuracy. In addition, the speed of the proposed system takes only 4.04 seconds, while the co-occurrence matrix method takes 24.81 seconds and the fast Fourier transformation method takes 33.83 seconds.