Active Sub-neighborhoods:Active Contour Model-Based Object Recognition Method

Region-based snakes deform according to the similarity in feature between the pixels on snake and the corresponding image regions, or the probability of the pixels on snake being in the corresponding regions, which can be formulized as a distance-like function. When the similarity or probability is measured, single pixel or one neighborhood data is adopted to evaluate the feature of the pixels on snake. However, using single pixel data makes snake sensitive to noise and texture, and using (whole) neighborhood information often result in an inaccurate boundary location. In this paper, we propose to use the curve itself to partition one whole neighborhood of a pixel on snake into two sub-neighborhoods, and then substitute the features evaluated by them for appropriate terms in the snake evolution equation. This very simple method retains the advantages but overcomes the disadvantages of only using single pixel or adopting whole neighborhood data. It improves region- based snakes in robustness to noise and texture, at the same time help them detect a more accurate boundary. It applies to almost all region-based active contour model, even if they are formulated as a level set form. Our presented experiments demonstrate the availability of the proposed method.