Research Of Shape Based Image Retrieval

Shape-based image retrieval is one important part of content-based image retrieval. Current shape-based image retrieval techneques mainly focus on shape feature extraction and feature similarity evaluation, combination of shape, color, texture features in image retrieval, and combination of shape feature with semantic feature in image retrieval. This paper analyzed some key technologies of shape-based retrieval, did some research work of shape retrieval based on Geometric Signature, Wavelet Scale Space and Wavelet-Fourier Descriptor, and improved some feature extracting methods.1. Improve the Geometric Signature shape feature extracting method. Normalize the shape feature to be rotation and scale invariant by using the information of the length and endpoints of the shape majoraxis, combine the feature vector acquiring with different majoraxis endpoints as the shape feature. The improved method overcomes the drawbacks of the Geometric Signature not being invariant to scale and rotation changes. Shape feature computed from combination of different majoraxis endpoints is more roubust to noise than the conventional rotation normalization methods.2. Improve the Wavelet Scale Space. Compute the improved shape feature with the length and endpoints information of the shape majoraxis. The improved method overcomes the drawbacks of the Wavelet Scale Space which is easily affected by changes of start point of shape boundary. This paper also studied the Wavelet-Fourier Descriptor.3. A testing system of shape-based image retrieval is completed with MATLAB and ACCESS, a user interface has been established; and according to MPEG-7 performance estimation scheme, 4 different image databases for testing performance of retrieval have been chosen and designed, and 3 retrieval performance estimation methods have been used to objectively estimate 8 different shape descriptors. The experimental results show that our shape descriptor has good retrieval performance, is more consistent with human perception.