A VISUAL SURFACE DEFECT DETECTION METHOD BASED ON LOW RANK AND SPARSE REPRESENTATION

Surface defect detection is very crucial for product quality control. A visual detection method based on low rank and sparse representation for surface defect detection of the wind turbine blade is brought forward in this paper. Two terms, which are the Laplacian regularization term and the noise term, were added into robust principal component analysis (RPCA). The noise term defined by F-norm is used to suppress uneven illumination and Gaussian noise, and the Laplacian regularization term is utilized to constrain the spatial relationship of superpixels. The defect image is considered to consist of a low rank matrix, a sparse matrix and a noise matrix, which corresponds with non-defect portion, defect portion and the noise portion of the image. At first, the proposed method segments the input image into a number of non-overlapping superpixels and extracts their features. Then, the optimal salient map is generated via the proposed method. Finally, the binary image is obtained by Otsu method. By quantitative and qualitative evaluation, experimental results illustrate that the proposed method is superior in terms of robustness and accuracy compared with 10 state-of-the-art methods on the synthetic and real images.