Improved multivariate image analysis for product quality monitoring

Abstract

Product quality monitoring by image texture analysis underwent a tremendous growth in the last few years in several industrial sectors, due to the availability of low-cost digital imaging sensors. Multivariate image analysis (MIA) can be used within an image texture analysis technique to provide a spatial statistical characterization of an image. However, in most cases this spatial characterization is possible only for very local texture neighborhoods, due to the high computational cost of MIA. In this paper, the iMIA (improved Multivariate Image Analysis) algorithm is proposed, that improves over previous implementations of MIA by extending its range of applicability due to its reduced computational complexity and memory requirements. The proposed algorithm uses the Fourier transform and the convolution theorem to efficiently compute the MIA model, in such a way that the image texture can be characterized by taking into account also large neighborhood sizes. The proposed approach is applied to two case studies concerning the estimation of the fiber diameter distribution in nanostructured membranes, and the classification of paper surfaces. The results suggest that the optimum range of spatial statistics used for characterizing the image is related to the size of the main textural features.