Defect Inspection of Patterned Thin-Film Ceramic Light-Emitting Diode Substrate Using a Fast Randomized Principal Component Analysis

Abstract

Manual inspection is commonly used to maintain the quality of light-emitting diode thin-film ceramic substrates. However, manual inspection has certain inherent and inevitable shortcomings. This paper describes an auto-inspection system for detecting defects on this type of integrated substrate surface using machine vision. A sequence of sub-images of an integrated substrate is grabbed and reconstructed into a high-resolution image. The unwrapped image is then shrunk to focus on the region of interest, which contains repeating and arranged substrates. A novel image restoration scheme based on randomized principal component analysis (RPCA) is proposed. That method is associated with the homogeneity cues difference histogram (HCDH) to construct a substrate defect map. RPCA is used to detect defective pixels while HCDH is used to estimate pattern size. The proposed method is compared with golden-template self-generating method, discrete Fourier transformation, non-negative matrix factorization and singular value decomposition for image restoration. The criteria for evaluating the different methods are the g-mean, figure of merit and average computation time. It finds that our method is more efficient and effective than others.