Ear recognition based on multi-scale features

This paper proposes a novel ear recognition method using multi-scale features inspired by the theory of the SIFT. Firstly, ear images are normalized by ear outer contour tracking and the longest axis detection. Then the difference of Gaussian (DOG) images are constructed using scale space theory and their corresponding block-based feature descriptors are determined. Finally we build the nearest neighbor classifiers and EMD is used as the dissimilarity measures. The weighted majority voting technique is used for decision fusion. Compared with other widely used ear recognition methods, such as PCA and KPCA, our method needn't transform the image to the same size and it is more robust to pose and illumination. Extensive experiments have performed to valid its efficiency.