Automatic detection of metal surface defects using multi-angle lighting multivariate image analysis

One of the key problems in detecting metal surface defects is that the lighting angles have great influence on the defect features information in the image. A multi-angle lighting multivariate image analysis approach was proposed to improve the accuracy and reliability of detection results. By adjusting the lighting height selectively, the surface images with multi-angle lighting could be obtained and used to constitute the multivariate images data, where each channel is the representation of the metal surface image with different lighting angles. It is based on the Multivariate Image Analysis (MIA) technique to extract defect features information. The effective lighting angles were selected according to score image and corresponding loading vector obtained by the Principal Component Analysis (PCA) strategy. Multivariate images with effective lighting angles were stacked and unfolded, from which the principal component scores of test images could be obtained. The Q-statistic image could be computed by removing first principal component score and the noise. With an appropriate threshold decided by training images and the morphological post-processing, the surface defects could be detected with accurate locations. Experimental work was performed. The results with lower pseudo reject rate verify the robustness and reliability of this method.