均值移動(Mean-shift)平滑技術於表面瑕疵檢測之探討

Abstract

Listen

Translate article

The mean shift technique has been an attractive alternative for noise removal, region segmentation and object tracking in image processing. In this thesis, the feasibility of mean shift smoothing for defect detection in complicated surfaces is studied. The proposed methods especially focus on low-contrast non-textured surfaces such as mura defects (uneven brightness) in TFT-LCD panels, and the heterogeneous surfaces such as polycrystalline solar wafers. Mean shift smoothing involves an iterative procedure that shifts each data point to the mode of the data points based on a kernel estimator of density. For non-textured surfaces, two mean shift-based methods are proposed. The first method shifts each pixel to the mode in the image, and the distance between the original pixel location and its converged position is used as the discrimination measure. A defect-free pixel will converge fast in its neighborhood and results in a small shift, while a defective pixel will need a larger shift to converge. In order to speed up the computation, a weight measure that uses the kernel function to calculate the gray-level variation in the spatial window in one single mean-shift iteration is also proposed for detecting low-contrast defects. For heterogeneous solar wafers, the fingerprint and contamination defects are studied. Since the grain edges in the polycrystalline wafer in a small spatial window show more consistent edge directions and a defect region presents high variation of edge directions, the entropy of gradient directions of each pixel in a small neighborhood window is first calculated to convert the gray-level image into an entropy image. The mean-shift smoothing procedure is then performed to remove defect-free grain edges in the entropy image. The preserved edge points in the resulting image are declared as defective ones. Experimental results have shown that mean-shift technique can be an effective tool for low-contrast defect detection in non-textured surfaces. It also performs well for defect detection in heterogeneous surfaces if the defect features can be adequately extracted.