| Summary: | 碩士 === 華梵大學 === 工業工程與經營資訊學系碩士班 === 94 === The purpose of this research is to apply image processing technology and neural network to the defective detection of array panel. Firstly, two-dimensional Fourier transform and inverse Fourier transform are utilized and then let defects be even conspicuous by the use of binary image segmentation technology. By the help of pixel labeling, defective breach detection and dilation and some other processing techniques to make the defective spots returned to the original condition. According to the image characteristics, we identify six graphic features: area, perimeter, roundness, length and width of the least rectangle covering whole spot, four-directional areas of defective spot, and four-directional areas of defective spot after thinning, that serve as feature values of that defective spot. All the six image characteristics are served as the characteristic values for that defected spot detection.
In conclusion, the feature values can be employed as input training and testing data for neural network paradigms. The recognition process can then conducted by the application of Self-Organizing Map (SOM) neural network, ant-based SOM, BPN and supervised SOM for three common metal residue defects, such as round residue, concentric residue, irregular residue, and scrape. Samples without defects are also added to do recognition. Experimental findings clearly show that these four neural network paradigms are capable of recognizing five kinds of defects in the array panel. The recognition correctness rate can achieve 100%. We may apply this research result to the online recognition system for defective detections of the array panel to reduce human detection errors.
|
|---|
