A Study of Defect Detection and Classification Technique for The Inverted-Mesa Crystal Images

• Main Authors: Hsin-Yung Liao, 廖信勇
• Other Authors: Jiunn-Lin Wu
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/33895317959084447854

碩士 === 國立中興大學 === 資訊科學與工程學系 === 102 === The inverted-mesa crystal which is etched in the crystal center to become a shape of trough can make quartz crystal oscillator frequency higher. Quartz oscillator is an electronic component to produce a high-precision oscillation frequency using a piezoelectric effect produced by the quartz crystal. It also can be called passive components. Inverted-mesa crystal oscillator is one of Quartz oscillator. The crystal defects will make frequency inaccurate or unstable. The crystal size is very small. The inverted-mesa crystal size in the thesis is 1.40mm * 0.75mm, and size of its defects had to be detected is 0.03mm * 0.03mm or more. Therefore, it is necessary to use a microscope for manual detection and it will result in high cost and low efficiency. So we hope to improve the shortcomings of manual defect detection by machine vision. Different types of defects may occur in the inverted-mesa crystal forming process. And the image color variation caused by light field variability and quartz glass refraction makes the analysis of defects difficult. There are many methods have been proposed to detect defects, but there is no one method can effectively detect all types of defects. The main difficulty is caused because of the color variation. If executed directly defects detection will spend a lot of algorithms and complex process and produce a high probability of miscarriage. In this thesis, we use color image as the framework to mainly overcome the sources of the problem, that is color variation. So we perform image pre-processing before detecting defect. The image pre-processing is color normalization[1]. It is non-defect inverted-mesa crystal color distribution as the standard, and normalizes the color distribution of the source image to the standard color distribution to eliminate color variation between images. And then we take out the region of interest to detect the defects. Due to the significant reduction of the source image color variation, so we can use fast binary methods to segment the defects and the threshold can be set in very small and very precise range. And then defect features extracted from the segmentation result are inputted to support vector machine (SVM)[8] to classify defects types. From the experimental results, this method can effectively eliminate the image color variation and normalize all acquired images to the same color distribution, and greatly reduce the difficulty and get high detection efficiency , low complexity and low false positive rate results.