Research Of Hand Vein Identification Based On Attention Mechanism

Hand vein recognition as an emerging research technology for biometric identification is of great significance for identity information verification and public security criminal investigation.In view of the subject's irresistible factors such as acquisition posture and collection environment,the recognition rate of the back vein image in the rotation,translation and brightness is reduced.This paper studies the attention mechanism of human vision and makes it fast.The ability to accurately locate significant areas is introduced into the dorsal vein recognition of the hand to complete an efficient and accurate hand vein recognition task.The main work is as follows:(1)In the case of a non-restricted subject's back posture,the subject is allowed to perform data collection in the most natural posture while constructing a multi-pose hand vein image database.In this database,the CNN-LSTM model and the Recurrent Attention Model(RAM)based on the attention mechanism are studied.Through experimental comparison,it is verified that RAM is more suitable for the identification of multi-attitude hand veins.The recognition rate is 89.69%.(2)The RAM extracts local features of the image of the dorsal vein of the hand based on the precise positioning of the salient points and the size of the significant area.Aiming at the problem that the position of discrete significant points in RAM and the scale of significant area lead to the decrease of recognition rate,this paper adopts the initialization point that is uniformly distributed by U(-1,1),and the model output obtains the prediction point.In order to make the prediction point more suitable for hand vein recognition,the Monte Carlo sampling method is used to obtain the random significant point It,and Gaussian noise is introduced to solve the position mutation problem of the random significant point.Focusing on this significant point,the significant area of the dorsal vein of the hand with the spatial resolution at the optimal scale is gradually reduced.Under this method,the recognition rate of the dorsal vein of the hand increased to 96.33%.(3)For the problem that the RAM loss function is not easy to converge,the baseline"b" is introduced in the reward value of the reinforcement learning,"b" is the output of the hidden layer node at different times,and the mean value error of the bonus value and "b" is calculated.Update the policy gradient function.At the same time,the regularization coefficient of the baseline "b" is introduced to the mixed loss function including the cross entropy loss function and the strategy gradient function to prevent the model from overfitting.Through theoretical analysis and experimental verification,a cyclic attention model with a node number is determined.The improved RAM has a recognition rate of 99.89%in the multi-anatomy hand vein.In summary,the recognition rate of the multi-anatomy hand vein increased from 89.69%to 99.89%,which proves the effectiveness of the work.