Using Neural Networks for Surface Defects Classification - A Pilot Study

• Main Authors: Chen, Chia Fa, 陳加發
• Other Authors: Gong, Dah Chuan
• Format: Others
• Language: zh-TW
• Published: 1994
• Online Access: http://ndltd.ncl.edu.tw/handle/50909392206083002781

碩士 === 中原大學 === 工業工程研究所 === 82 === The machine vision system apply in the production line is the solution for the dificulty of the automatic production process. The merchandize of the machine vision system is used statistics paptern recognition as classifier method. For statistics and structure ayalysis method need a complex analysis for each pattern of data. So, in this study, we will use neural network to classify the defects of the oil-lid. In this study, image substraction, image matrix transfer and neural network are employed to model the machine vision system. Adaptive Resonance Theory(ART) and Back-Pagation(BP) network are used which learning and training through five defect patterns and a good one images, then used such neural networks to classify the oil-lid which is a good production or one of the five kind defect of production. The result of the experiment is: First, without any refined method: CGNN:65%, ART1: 80%, ART2:87.3%, BP:90.67%. Second, used one refined method: CGNN:75%, BP:95%. Through refined method, the BP network has been proved that machine vision system has a good result in the pattern recognition. The major dificulties in refining process of the classfication system. First, the choice of refining system methods depend on the result of the experiment. Second, it would not know the reason after the result of refined is bad. Third, it would not know to stop refining under the unknow refined value. In this study, it is fucus to solve the troubles of using refined methods, we can make a bettre result for neural network in the pattern recognition.