Paimprint Recognition Based On Deep Learning

Informational interactions in modern society required high accurate personal authentication, and it gave rise to a biometric identification technology, public safety and commercial finance department. Biological characteristics that usually used for identification mainly included face, fingerprint, palmprint, vein, iris, gait, voice, signature, etc. Palmprint recognition, as one of many biological recognition technologies, got plenty of attention of many research team, with its abundant information, stable and unique features, simple collection devices, little affection from the noise, higher acceptance from users, etc. Common feature extraction of palmprint algorithms at present were shallow algorithm fundamentally, this article tried to extract feature in the deep method, and the main research contents were listed as follows:(1) Palmprint recognition based on deep learning was realized. Applied deep learning method which were considered as a breakthrough in the world of computer vision to palmprint recognition. Built a deep learning model after the image preprocessing, and adjusted the parameters such as the number of hidden units, the learning rate, to an optimal model, then, classified characteristics by Softmax. Deep learning algorithm realized complex function approximation through the studying of nonlinear network structure, with a strong ability to get the essence from a small amount of images. This article applied deep learning algorithm to palmprint recognition to get the essential characteristics through unsupervised training and supervised training. Deep learning method showed a good performance with the database.(2) A new method of palmprint recognition based on lift wavelet and deep learning was proposed. If the original images were put into DBN directly, in the form of a one dimensional vector, the model would be difficult to give attention to the local feature details of the images, what's more, undesirable characteristics would be learned for adverse factors such as illumination, tilt, which ultimately affected the recognition results. Considering the shortcomings of deep learning model, a new method in which, palmprint images were wavelet processed to get the local detail information, and a designed DBN network was used to automatically extract more effective features. In this way, advantage of the lift wavelet was fully taken and the defect of the deep learning model was maken up. The experimental results demonstrated better performance of the proposed algorithm compared with traditional algorithms, such as LBP, PCA, and original DBN based on pixel level. Therefore, the re-extraction of DBN based on the initial characteristics gained by lift wavelet could effectively obtain the robust features of palmprint images.(3) A new method of palmprint recognition based on image reconstruction and double DBNs was proposed, as principal component analysis algorithm ignored the higher order statistics in feature extraction. First whitening PCA method was utilized to extract the characteristics of the original palmprint image, the original image were reconstructed, the residual image were calculated and the characteristics of the residual image were extract in the same way, then the a new double DBNs were used to secondary feature extraction and classification prediction. The experimental results demonstrated better performance of the proposed algorithm compared to traditional algorithm (PCA, LBP, HOG) and DBN algorithm based on pixel level, and got higher recognition rate, verified the effectiveness of the algorithm.