For the applications of face recognition, a learning based kernel image differential filter was proposed. Firstly, instead of designing the image filter in a handcrafted or analytical way, the new image filter was designed by dynamically learning from the training data. By integrating the idea of Linear Discriminant Analysis (LDA) into filter learning, the intra-class difference of filtered image was attenuated and the inter-class difference was amplified. Secondly, the second order derivative operator and kernel trick were introduced to better extract the image detail information and cope with the nonlinear feature space problem. As a result, the filter is adaptive and more discriminative feature description can be obtained. The proposed algorithm was experimented on AR and ORL face database and compared with linearly learning image filter named IFL, kernel image filter without differential information, and kernel image filter considering only one order differential information. The experimental results validate the effectiveness of the proposed method.

The algorithm of Fast REtinA Keypoint (FREAK) descriptor has achieved the rotation invariance via the direction of calculation model, but its matching performance for large change of rotation scale is not ideal and the matching error rate is high. In order to solve the problem, an improved image registration algorithm based on FREAK descriptor was proposed. Firstly, long distance point pairs judged with a given distance threshold, was added to the original FREAK. Only the points of long distance in the keypoint sampling pattern were used to generate angle information. Then, the Hamming distance was weighted. In order to generate descriptor selection point pairs for every key point, the mean of each column of training data descriptors was computed. The mean was closer to 0.5, the weight of the column was larger. This method improved the coarse-calculating state of original Hamming distance and made the distance calculation more accurate. The nearest neighbor matching method combined with the ratio of the nearest neighbor and next nearest neighbor, and the method of RANdom SAmple Consensus (RANSAC) were used for rapid matching and optimization. The experimental results show that, the improved algorithm is more suitable for the applications with large variation of rotation scale and high demand of matching performance.