Multiscale geometric feature extraction and selection algorithms of similar objects

To recognize objects with similar shapes, a scheme for feature extraction and selection based on Multiscale transformation is proposed in this paper. Multiscale Geometric Analysis is characterized with directionality and anisotropy, and the subbands in different decomposed scales could present different classification capabilities. The scheme applies time-frequency-localized feature algorithm as well as probability information measurement to choose the decomposing scale and directional subband in order to maximize similarity between objects in the same class while minimize similarity of objects in different classes. To some extent, the algorithm proposed has resolved the random selection problems of decomposing scale, direction number and directional sub-bands in Multiscale transforms. The experimental results have verified the effectiveness of the algorithm.