Transforming pixel signatures into an improved metric space

We address the problem of using scale-orientation pixel signatures to characterise local structure in X-ray mammograms, though the method we report is of general application. When signatures are treated as vectors for statistical analysis, the Euclidean metric is not well behaved. We have previously described a Best Partial Match (BPM) metric that measures signature similarity more naturally, but at high computational cost. We present a method for transforming signatures into a new space in which Euclidean distance approximates BPM distance, allowing BPM distance to be estimated at low computational cost. The new space is constructed using multi-dimensional scaling. The nonlinear transformation between the old and new spaces is learned using support vector regression. We present experimental results for mammographic data.