Relevant shape contour snippet extraction with metadata supported hidden Markov models

In most existing works on shape contour matching, the shape contours are considered and matched in whole. When searching for contour snippets, however, techniques that match whole contours are not directly applicable. In particular, a relevant snippet can be anywhere on a shape contour; moreover, the relevance of shape snippet is a function of not only the shape of the snippet itself, but also of its neighborhood on the contour. In this paper, we propose an HMM based solution to shape snippet extraction. Relying on a general-purpose symbolic representation (such as SAX), we first convert the shape contour onto a representation suitable for snippet marking and extraction processes. We then show that, given a set of samples, we can train an HMM capable of detecting relevant snippets in new shape images. Next, we show that the HMM performance can be boosted significantly if the similarities between the symbolic representations are used to create new sibling training sequences from the input sequences. The experiment results show that just adding one additional sibling per input training sequence can improve the diversity of the training set sufficiently to boost the overlaps between actual and detected snippets much. While a naive application of this metadata driven training technique can increase the training costs significantly, we show that a novel metadata-driven HMM (mHMM) scheme can significantly improve the HMM-base snippet detection performance with negligible costs.