Research On Convolutional Neural Networks-based Finger Vein Descriptor And Its Effectiveness

Biometric technology uses the inherent physiological or behavioral characteristics of the human body to identify personal identities,and has been widely used in finance,education,medical care,social security and other fields.Among them,the finger vein recognition technology obtains the vein image of the finger through the infrared CCD camera,and extracts finger vein related features for identity authentication,which has the advantages of privacy,security and convenience.However,during the finger vein image acquisition process,due to the interference of the external environment,the quality of the finger vein image is often low;in addition,because the finger is not fixed,it is easy to cause posture changes such as translation and rotation.In the face of these problems,especially the changes of finger posture of the same identity,the traditional finger vein descriptor extraction method is usually difficult to obtain a robust descriptor which often leads to poor recognition accuracy.In this paper,we based on convolutional neural network to carry out related researches.Firstly,we proposed a framework which combines feature description with feature encoding to get robust finger vein descriptor.Then,we explored more effective methods from two aspects of network structure and loss function.Experiments show that the proposed method achieves the best finger vein recognition results on SDUMLA,FV-USM and MMCBNU data sets,and has theoretical significance and practical value.The specific research work is as follows:1.A finger vein descriptor extraction framework(Describing and Encoding Framework,DAE)combining feature description with feature encoding is proposed,which mainly includes a preprocessing module,a CNN-based local descriptors extraction module and a descriptors encoding module.The proposed framework introduces the VLAD coding method and is trained by triplets,which make the descriptor more discriminative and robust to the changes of finger posture.When using the same basic network,the descriptors extracted by the DAE framework have better robustness against random translation.2.Based on the DAE framework,two lightweight finger vein descriptor extraction networks are proposed.By reducing the feature map channels and the number of convolution layers of VGGFace-Net,it is more suitable for finger vein images with fewer features and smaller size.We use the reduced network as a local descriptor extraction module for DAE to get a model with parameter size of 0.3M,which improved the accuracy of recognition and verification;Further,by adding feature map reuse method and 1×1 convolutional layer to the network,we reduced the extracted descriptor dimension from 1280 to 640 which basically maintains the recognition accuracy while improving the matching speed.3.A Pair-center-constrained Loss is proposed for training finger vein descriptor extraction network.Since there is no constraint on the within-class variance of the sample pair distance in Triplet Loss.By considering the positive and negative sample pairs as two categories,constrain the positive and negative sample pairs to be close to their centers,so that it is easier to find a suitable binary classification threshold to divide the positive and negative samples.Experiments show that using the proposed loss function has a lower error rate in finger vein verification experiments than the Triplet Loss.