Research On Hand Vein Recognition Based On Deep Learning

Entering the21th century, along with the high-speed development of information technology, particularly the quick development of Internet, information security problem become more and more prominent. People began to care about how to ensure their information security. Biometric identification technology is also a kind of safe and convenient technology and develops rapidly. There are now many biometric identification methods, for example, fingerprint recognition, face recognition, red membrane technology and so on. Biological recognition technology has been widely used in the field of passport, management, access control, financial and so on. Hand vein recognition technology is also a kind of biological recognition technology, which is safe, convenient and friendly. Hand vein recognition technology gradually draws attention of the scientific research institutions. So far, some products of hand vein recognition have been made and applied to usual life.The research is concentrated on the achievement of hand vein images, correction of hand vein images and ROI extraction, LBP feature extraction of hand vein images, deep learning training, code combination. First of all, hand vein images of more than one hundred of persons are obtained with hand vein image acquire device, and are put into one database. By using knuckles information of dorsal vein images, the angular deviation of hand vein images are corrected. After angular deviation being corrected, region of interest of the hand vein images are extracted. In the research, LBP was used to extract the texture features of hand vein image. Secondly, deep learning method is adopted to train the LBP features of different images. Traditional BP neural network is one of the effective ways for classification, But it is hard to get global parameters when optimizing the weights, Because BP network should be given initial weights for optimization. If the weights are large, it is easy to converge to the local bad point. If the weights are small, the weights’ update of layer close to input layer is slow when making error reverse transmission. But deep learning network can find a better global optimal point. Compared to BP, one of the most of important breakthrough is that it conquers the problem of training efficiency and training effect of shallow neural network. In the frame proposed by Hinton, the strategy of deep learning is that multilayer networks are divided into several RBMs’superposition and trained step by step before learning the global parameter of multilayer model. In the research, to reduce the rate of error, correlation classifier method is adopted.