Improved image features by training non-linear diabolo networks

This paper discusses a trainable system to extract features for image segmentation based on non-linear mapping of local features. Supervised training methods are presented, for artificial neural diabolo networks, which produce a mapping comparable to Fisher’s linear discriminant mapping. This mapping can be used to decrease dimensionality whilst preserving class separability. It is shown that the non-linear feature extraction performed in diabolo networks can increase class separability, compared to linear mapping methods, thus resulting in improved image segmentation.