Multi-Scale Feature for Recognition

For combining global and local features effectively, a multi-scale description and feature extraction algorithm is proposed. The original image is decomposed into two levels by wavelet analysis, and the two reconstructed approximate images are divided into several regions. On each region the singular value features are extracted, and then these singular value features are organized and used as an eigenvector of the image. Finally Fisher linear discriminant analysis is used for classification and recognition under these multi-scale singular value vectors. The experiments were made on ORL face database with recognition rate of 97.5%, and on ear database with recognition rate of 98.33%. Compared with corresponding algorithms, the proposed algorithm can achieve high recognition rate under the low dimension eigenvector. The results show that the multi-scale singular value vector includes not only global feature but also local feature of image, so more discriminant information for pattern recognition is contained.