Defect detection in periodically patterned surfaces using independent component analysis

In this paper, we propose a fast self-comparison scheme for defect detection in structural surfaces containing periodic complicated patterns. It works directly on a one-dimensional line image, instead of a two-dimensional array image, that contains a periodic pattern in the line. The proposed self-comparison scheme is simply carried out by dividing a sensed line image into two segments of equal length. Since the line image contains a periodic pattern, the two divided segments are only translated versions to each other. In this study, an independent component analysis (ICA) model is proposed to obtain the de-mixing matrix that can recover the translation between the two divided segments. The proposed ICA model directly measures the independency of signals by minimizing the difference between the joint probability density function (PDF) and the product of marginal PDFs, in which the PDFs are estimated by relative frequency distributions. The particle swarm optimization (PSO) algorithm is used to search for the de-mixing matrix. The proposed ICA model can effectively separate highly correlated signals, and is well suited for translation recovery between two signals with the same periodic pattern. In the detection stage, each line image is first divided into two segments, and the de-mixing matrix teamed off-line from a defect-free line image is used to recover the signals with well aligned translation. The normalized cross-correlation is adopted to measure the similarity between two compared segments. Since the de-mixing matrix is only of a small size of 2 x 2, the proposed method in the detection stage is very computationally efficient. The performance of the proposed method is demonstrated with test samples of TFT-LCD panels and color filters found in LCD manufacturing. Experimental results have shown that the proposed self-comparison scheme can effectively and efficiently detect the presence of defects in periodically patterned surfaces. (c) 2008 Elsevier Ltd. All rights reserved.