Texture analysis of images using Principal Component Analysis

Extracting texture/roughness information from grayscale or multispectral images for off-line quality control, or on-line feedback control is a difficult problem. Several statistical, structural & spectral texture analysis approaches for grayscale images (using various pre-defined filters etc.) have been suggested in the literature(1,2.) In this paper we propose a new approach based on Multivariate Image Analysis techniques using multi-way Principal Component Analysis. Prior to analysis the grayscale images are transformed into three-dimensional pixel intensity arrays through spatial shifting of the image in several directions followed by stacking the shifted images on top of each other. The resulting three-dimensional image data is a multivariate image where the third (i.e. variable) dimension is the spatial shifting index. Multi-way PCA is then used to extract features (PC scores), which contain the greatest amount of variation. Plots of the observed values of these scores against one another define a score space. Certain regions of this score space contain the texture information of the grayscale image. By masking these regions and tracking the number of pixels having features that fall in these regions, or by comparing the score spaces with template exemplars, one is able to monitor changes in the image surface textural properties. The approach is illustrated using a set of grayscale images of the surface of steel sheet. Based on the textural features extracted from the surface images a simple classification scheme is devised in which each sample image is assigned into one of two classes representing good or bad surface characteristics.