Research On Finger Vein Recognition Method Based On Multi-Scale Gabor Filtering And Attention Mechanism

In recent years,finger vein recognition technology has become a research hotspot in the field of biometric identification.It has the advantages of live capture,not easy to forge,portable collection,and accurate identification.It is widely used in identification,information security and other fields.Although the finger vein recognition algorithm has made certain research progress,in practical applications,finger vein images are of low quality and are easily affected by rotation and translation.Therefore,it is still challenging to establish a robust finger vein recognition system.Generally,finger vein recognition methods include four steps: finger vein image acquisition,preprocessing and region of interest extraction,feature extraction,and matching recognition.Feature extraction is a key step,which directly affects the accuracy of subsequent recognition.This dissertation mainly studies the method of finger vein image feature extraction,which mainly includes the following contents:(1)Finger vein recognition based on multi-scale and multi-directional Gabor filter.The original Weber local descriptor,the method of calculating the differential excitation and the gradient direction is difficult to effectively represent the characteristics of the finger veins.This dissertation first uses a series of multi-scale and multi-directional Gabor filters to filter finger vein images to obtain multi-scale energy maps and directional maps;secondly,calculate the differential excitation of the pictures based on the energy maps,and select the largest and second largest responses after Gabor filtering.Finally,the joint statistical distribution feature is constructed based on the multi-scale differential excitation map and the dual Gabor pattern,and the feature vectors obtained after Gabor filtering of different scales are connected in series to obtain the finger vein feature vector.Experimental results show that the proposed algorithm can overcome the influence of rotation and scale to a certain extent.(2)Finger vein recognition based on SE(Squeeze-and-Excitation)attention mechanism.Different from the traditional finger vein recognition algorithm research,the convolutional neural network improves the distinguishing ability of the feature by self-learning features.This dissertation first stacks the Inception-resnet module to build a main body neural network to extract features from finger vein images.This network incorporates multi-scale processes and has better generalization learning capabilities for features;secondly,the SE structure is used to upgrade the network structure,and the original network is introduced the attention mechanism enables it to use the correlation of the global information channel to strengthen the important features while reducing the complexity of the model and improving the recognition accuracy of the vein network.Finally,the network model is trained to obtain the optimal parameters of each layer of the network for experiments,and the finger vein feature vector is obtained.The experiment was carried out on three public databases.The results proved that compared with the traditional finger vein recognition algorithm,the method in this dissertation can better solve the problems of sample rotation and translation in finger vein recognition,and improve the recognition performance to a certain extent.