Research On The Key Algorithms For Palmprint Recognition

The palmprint is known as all the lines contained in the inner surface of a palm, including principal lines, wrinkles and minutiae, etc. These features are stable and unique, which makes palmprint become an important biometric feature for personal identification and verification. Meanwhile, with the obvious advantage of low cost, low resolution and non-intrusiveness, palmprint recognition has attracted wide attention from the researchers.Extensive foundations of theory and application have been made in the field of palmprint recognition over ten years' research. Especially for the preprocessing and segmentation stage, the algorithms are relatively ample and mature. However, the algorithms for recognition are much weaker. The sophisticated features of palmprint and variations of the images such as translation, rotation, distortion, and the delta regions formed by palm structure, affect the recognition performance of the algorithms seriously. Therefore, to solve the above issues, the dissertation proposes several algorithms with high recognition rate and steady performance from three aspects: subspace analysis, feature fusion and palmprint texture features, on the basis of the extensive research the existing algorithms for palmprint recognition. The main contributions are listed as follows:1) An improved two-dimensional locality preserving projections (I2DLPP) is proposed for palmprint recognition. Different from the traditional appearance-based subspace methods, 2DLPP emphasizes on keeping the manifold structure of image space. Whereas, the nearest neighbor graph based the image as integration fails to model the intrinsic manifold structure inside the image. Therefore, the nearest-neighbor graph is proposed in which each node corresponds to a column inside the matrix, instead of the whole image, to better model the intrinsic manifold structure. In addition, the proposed I2DLPP implements two-directional projections of 2DPCA and 2DLPP to reduce the calculation complexity and final feature dimension. The experimental results demonstrate the effectiveness of I2DLPP in recognition accuracy and speed. 2) A fusion algorithm of global and local features for palmprint recognition is proposed. The global features can reflect the main information of a palmprint image, and the local features are the most discriminant in recognition. The fusion of both features can overcome the locality and uneven impact on the recognition performance. The frequency global features are extracted by the proposed algorithm Gabor-based (2D)~2PCA(GB(2D)~2PCA). The algorithm is to form the Gabor space of high dimensionality with relatively complete frequency features extracted by multiple Gabor filters, and then use (2D)~2PCA to reduce the dimensionality in two directions to extract the most representative features to strengthen the robustness against the variations of the circumstance. Compared with the single algorithms of Gabor and some subspace algorithms, the GB(2D)~2PCA improves the recognition performance in accuracy and efficiency.The spatial local features are extracted by NMFSc. The fusion procedure of global and local features is to normalize the match distances of GB(2D)~2PCA and NMFSc by zero-mean normalization, and then fuse them with the weighting sum strategy to get the final matching distance for recognition. Owing to the strong supplementary of the global features and local features, the proposed fusion algorithm gains a much higher correct recognition rate. Meanwhile, the two subspace algorithms are easy to execute with a better recognition performance. The simple fusion strategy without extra training time has no effect on recognition speed.3) Gabor local relative features are proposed for palmprint recognition. The parameters of Gabor filters involve scales and orientations, which guarantees the invariance of features. After two-layer partitions of the Gabor filtered image, local relative variance can be defined as the relationship between lower-layer range block and its resided upper-layer domain block. The feature vector to represent image is composed by the local relative variance of all the lower-layer range blocks. Owing to the invariance of Gabor features and relatively stable features, the global disturbance occurred on palmprint images is counteracted. At the same time, the proposed method is of high efficiency for the relative features are simple statistical quantities without sophisticated calculation and Manhattan distance is used as the similarity measurement in the procedure of feature matching.4) Contourlets-based local fractal dimensions (CLFD) is proposed for palmprint recognition. The novelty of CLFD comes from using fractal dimension independent of time-consuming fractal coding for palmprint image representation, which can meet the requirements of real time system. In addition, Contourlets can extract the sophisticated texture of palmprint, with the anisotropy of the multiple directions and scales and local time-frequency characteristics. Therefore, the method is robust to the variations and distortions occurred on palmprint images, resulting in higher recognition accuracy. The efficiency of CLFD can be guaranteed by using DBC to calculate the fractal dimensions of local blocks and the Manhattan distance as the similarity measurement.