Optimized multi-scale affine shape registration based on an unsupervised Bayesian classification

Khaoula Sakrani, Sinda Elghoul, Faouzi Ghorbel

Abstract: Here, we intend to introduce an efficient, robust curve alignment algorithm with respect to the group of special affine transformations of the plane denoted by SA(2,R). Such a group of transformations is known to be well model the pose of 3D scene when objects are far from the visual sensor relatively to their seizes. Its numerical robustness lies in its multi-scale approach and its precision comes from the automatic and unsupervised Bayesian selection of the efficient scales in the sens of L2 metric. In this work, We prove its high alignment performance on the most studied image databases such as MPEG-7, MCD, Kimia-99, Kimia216, ETH-80, and the Swedish leaf experimentally. The unsupervised Bayesian classification is based on the well-known multiclass Expectation-Maximization algorithm.

External IDs:dblp:journals/mta/SakraniEG24