Fast Shape Retrieval Based on Invariant Feature Alignment

This paper proposes an invariant feature alignment for shape-based image indexing and retrieval. We have found that some recently published systems have employed complex shape matchers to achieve high accuracy, at the cost of relatively high pairwise matching times. Using invariant feature alignment, our method can effectively increase the speed of shape retrieval without sacrificing accuracy. The key to its success is the use of a normalization procedure to generate a linear number of canonical reference frames as a function of the number of key points, the use of Haar transform to fast decompose the normalized curse into approximate polygons, and the use of fixed correspondences on those Haar expansion coefficients. The experimental results and comparisons show the good performance of the proposed technique