Bayesian labelling of corners using a grey-level corner image model

In this paper we present a recasting of corner detection to a problem in statistical pattern recognition which we then address with a simple feed forward neural network. The resulting classifier is a robust, threshold-free corner detector which labels with (approximate) Bayesian posterior probabilities; this is in contrast to conventional feature detectors which produce binary labels contingent on a heuristically set threshold. We have generated the training data for our classifier using a grey-level model of the corner feature which permits sampling of the pattern space at arbitrary density as well as providing a validation set to assess classifier generalisation. Results are presented for real images and the robustness illustrated over a well-known state-of-the-art conventional corner detector.