A new robust region-based ICA-SIFT shape descriptor for object recognition

摘要

The paper proposed a new region-based ICA-SIFT shape descriptor. It combines two methods to realize the optimal performance: ICA to process global information and SIFT to get local features, therefore, it can describe various kinds of shapes accurately and concisely. The ICA-SIFT shape descriptor is proved to be invariant to skewing, scaling, translation and rotation. The main process of the ICA-SIFTSD is first to extract the canonical form of an original shape by ICA and has eliminated any effects of skewing and affine transformation. Next, we carried out SIFT feature extraction on canonical forms and got the ICA-SIFT shape descriptor, which is an improvement of the ICAZMSD method in [3]. We have applied FastICA and SURF( the speedup of SIFT) that can accelerate the calculation speed of the proposed method so that it can meet requirements of real-time applications. In the paper, we carried out a large number of experiments on the MPEG-7-CE database, using the ICA-SIFTSD as an effective descriptor for object recognition. The experimental results show recognition rates are 91.7% and 93.8% of simple and complex shape images respectively.