| Summary: | 碩士 === 元智大學 === 工業工程與管理學系 === 90 === In this research, novel similarity measures are presented for automated defect inspection. Traditional normalized correlation approach has been extensively used as a similarity measure for pattern matching. However, it cannot provide good discrimination for detecting subtle defects in complicated images. The purpose of this study focuses on finding effective similarity measures, especially for defect detection applications.
The core idea of this study comes from conceptually constructing a gray-level corresponding map for two compared images. The x-axis and y-axis of the corresponding map are defined by the gray values of the reference image and the scene image, respectively. The pair-wise gray levels of each pixel coordinates in the images form a diagonal straight line in the corresponding map if the two compared images are identical. Any two compared images different to some extent will not have the shape of a line in the map. Eigenvalues and major-axis angle of the covariance matrix of the data points in the map are used as similarity measures to evaluate the difference between two compared images.
The proposed eignevalue-based similarity measures have better discrimination capability, and are more stable for defect detection application, compared to the normalized correlation. Experimental results on real industrial samples such as PCB, SMT, and printed characters have shown the efficacy of the proposed similarity measures.
|
|---|
