The Mahalanobis classifier with the generalized inverse approach for automated image processing

Two problems associated with image data processing are identified: multicollinearity and the nonnormal distribution of spectral/texture data. With reference to the normality problem, numerical evidence shows that the stable distribution may be used to develop a classifier using nonlinear discriminant functions. This would lead to a high hit-rate requiring reduced processing time. Multicollinearity may be solved by employing a generalized inverse technique in computing the Mahalanobis distance. In this way, unpooled individual dispersion matrices are used in the classification process. An improvement of eight percentage points in the classification rate has been observed.