Automatic masking in multivariate image analysis using support vector machines

A new masking method for Multivariate Image Analysis (MIA) is proposed. By interpreting the masking process in MIA as a classification problem it is possible to find masks automatically using any classification method and it is shown that Support Vector Machines (SVMs) are a good candidate for this purpose. An easy to use, iterative mask-building scheme, based on SVMs, is presented and used to find multidimensional masks for selected features in the Near Infra Red (NIR) imaging of lumber. Although the automated masking procedure can be used to obtain better masks in two-dimensional score spaces, typically obtained from RGB images, the real value of the procedure lies in its ability to easily obtain better-performing multidimensional masks from hyperspectral images.