Research On Recognition Technology Of High Resolution Palmprint Image Based On Deep Learning

Palmprint recognition is an important member of biometric recognition.High-resolution palmprint images refer specifically to palmprint images with a resolution above 300 dpi.The rich information contained in the image can be used for identification.Matching of minutiae features is the mainstream high-resolution palmprint image recognition method at present.Matching of minutiae features algorithm is difficult to achieve civilian and commercial use because its algorithm design is complex and it is difficult to manually extract features.In response to the above problems,this paper proposes two high-resolution palmprint image recognition methods based on deep learning.After simple preprocessing of the palmprint image,end-to-end recognition can be realized,which has the advantages of high recognition accuracy,easy implementation,and strong robustness.The main research results of this article are as follows:In view of the small sample data set of high-resolution palmprint images,this paper takes the capsule network as the research object.In order to provide better quality capsules for the routing algorithm,the ultra-deep small convolutional neural network is used to optimize the feature extraction part.By removing the reconstruction network in exchange for the simplification of the model volume and the improvement of the calculation speed,the algorithm complexity is greatly reduced under the limited accuracy loss.The routing algorithm is sensitive to the location of palmprint features,but could not distinguish the importance of different features,a channel attention mechanism is added in front of the main capsule layer to increase the weight of important features and further improve the recognition ability.Experiments on the THU high-resolution palmprint data set prove that the capsule network based on reasonable feature extraction has good palmprint recognition accuracy.The traditional convolutional neural network vgg16 network uses a deep small convolution kernel to extract features,which has better feature extraction capabilities and less computational overhead.Based on the vgg16 network,this paper proposes a method based on transfer learning,which can learn parameter information of similar features from the source domain,so that the network does not have to be trained from scratch.In addition,add a fully connected layer behind the vgg16 network to improve the classification ability of the network.On the basis of the vgg16 transfer network,image enhancement technology is used to divide the high-resolution palmprint images into 4,9,16,25 equal parts,and the accuracy rate of the palmprint image obtained by the voting method can reach 99.69%.