The success of automatic pattern recognition systems depends on the enhancement of significant features in relation to the irrelevant background information in the pre-processing stage. In images, the objects of relevance are generally enhanced by linear/non-linear stretching, histogram equalization and spatial filtering, which are all operations in a single band (univariate). In a multivariate space, linear transformations such as principal component analysis are very popular for this purpose. A simple rotation of axes, as in the principal component transformation, to the maximum variance direction is often insufficient to enhance objects camouflaged by the background. This is often due to the enhancement of the background together with the features of interest or non-background. In this paper a new technique is presented to address this problem. The images are modelled as having two classes, namely background and non-background. The technique, called background discriminant transformation (BDT), is designed to maximize the non-background class variance relative to the background variance. The technique has applications to image enhancement in mineral exploration, planetary sciences, biological and medical sciences, and defence applications.