Content-Based Superpixel Matching Using Spatially Constrained Student’s-t Mixture Model and Scale-Invariant Key-Superpixels

This paper addresses an image matching methodology designed for correspondence problem in computer vision. Firstly, a novel superpixel segmentation model driven by spatially constrained Student's-t mixture model (SMM) is proposed. The tails of Student's t-distribution are heavier than that of traditional Gaussian distribution, therefore, SMM is more insensitive to outliers and noise. In this model, a spatially constraint term based on Markov random field (MRF) is designed, so that good boundary adherence and intensity homogeneity would be achieved. Next, by constructing an adaptive superpixel Gaussian filter and a superpixel salient detector, this paper establishes an innovative key-superpixel detection method by building a superpixel scale-space pyramid. Different from conventional keypoint based detection, two images could then be matched directly in a superpixel-to-superpixel manner. During the matching process, a combinatorial feature descriptor that merges color, shape, gradient and texture features is set up to distinguish each considered key-superpixel. One main advantage of this approach is that implementation time would be largely reduced by less matching demand for key-superpixels and few corresponding local features. Some experiments on datasets at the end would demonstrate a relatively better performance of our model.