A Hybrid top-down/bottom-up approach for image segmentation incorporating colour and texture with prior shape knowledge

Image segmentation is maybe one of the most fundamental topics in image processing. Among numerous methods for segmentation, blind (bottom-up) algorithms, which are based on intrinsic image features, e.g. intensity, colour and texture, have been used extensively. However, there are some situations such as poor image contrast, noise, and also occlusion that result in failure for blind segmentation methods. Therefore, prior knowledge of the object of interest must be involved in the segmentation approaches. For this purpose, in this work, a novel integrated algorithm is proposed, which is a combination of bottom-up (blind) and top-down (including shape prior) segmentation algorithms. In our approach, after a colour space transformation, an energy function based on non-linear diffusion of colour components and directional derivatives, is defined. After that, some distance maps of the object of interest are generated from binary images that contain the training shapes of the object. Finally, the energy function minimization is done by evolving a level set function, which is set up by the distance maps. The results show our region-based segmentation algorithm robustness against noise and occlusion.