Optical-flow-based template matching for surface defect detection

• Main Authors: I-Yung Chiang, 江宜勇
• Other Authors: Du-Ming Tsai
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/32667123151452728602

碩士 === 元智大學 === 工業工程與管理學系 === 98 === Template matching has been widely used in image processing for visual inspection. The current existing methods such as normalized cross correlation and golden template matching are very sensitive to displacement even the inspection object is well aligned with its fiducials. Some object surfaces found in industry may show repeated patterns or contain no textures and, therefore, there are no fiducials can be uniquely chosen for alignment. This research proposes an optical-flow-based template matching method for surface defect detection. Given a reference image, the optical flow field between the reference and the inspection image are evaluated. The optical flow of each pixel is calculated within a small neighborhood window. When the pixel value of the inspection image is similar to that of the reference image, the flow length will be short. It indicates the pixel under evaluation is defect-free. Conversely, the pixel value of a defect in the inspection image is distinct from that of the reference image and, thus, the corresponding flow length is significantly large. The flow length derived from the optical flow field is insensitive to minor shifts of an inspection image. It is then robust to displacement due to the misalignment of the algorithm or random production variations. The optical flow process is computationally expensive. The integral image technique is applied to replace the sum operations in a rectangular neighborhood window, and its computation time is invariant to the window size. In this study, the optical flow algorithms for defect detection in both gray-level and color images are individually proposed. The experiment on LED wafer inspection that involves needle mark shift defect-free samples and large area defective sample has shown that the proposed method can achieve 100% recognition rate. The computation time of an LED chip of image size 115×105 pixels takes only 0.0121 seconds with the help of integral images. Additional experiments on structurally-textured surfaces such as textile fabrics and non-textures surfaces such as LCD glass substrates have also shown that the proposed method outperforms the conventional template matching methods.