Research On Finger Vein Recognition Algorithm Based On Deep Learning Network

Among the many biometric traits,finger vein trait has attracted people's interest due to its high security and is widely used in finance and security.However,different finger-vein acquisition devices have different methods,light source wavelengths,etc.,resulting in significant differences in background noise and the gray level of the images,which makes it challenging to locate the finger area.Also,the quality of the finger-vein images is often poor,and there are posture changes such as finger plane rotation and space rotation.Although deep learning-based methods can better extract robust features,the public finger vein database has a small scale,which limits the performance of the deep model.And as the application of the finger vein system has become more and more extensive,the types and number of finger veins have rapidly increased,and the fast retrieval of finger vein images has also become a research hotspot.We have researched these issues.The work has theoretical significance and practical value.The main research work is as follows:1.We propose a watershed-based finger vein region of interest(ROI)extraction algorithm.We first use the watershed method to perform superpixel segmentation on the image.The segmented superpixel boundaries are robust to disturbances such as grayscale changes and background noise.Then we define the tracking rule to extract the edge line of the finger and correct it according to the deflection angle of the midline line.Finally,we intercept the circumscribed rectangle of the finger and normalize it as the ROI image.Experiments show that the algorithm can accurately extract ROI from vein images collected from different devices.2.We propose a finger vein recognition algorithm that combines multiple schemes.We first introduce the lightweight neural network Shuffle Net V2 to construct a basic model.And then,for the characteristics of less training data and small ROI size,we start from the input end,network end,and loss function of the model to combine multiple schemes to improve the effectiveness of the model.These schemes include data augmentation,enlarge the output feature map,transfer learning,label smoothing,and joint loss function.Experiments show that these schemes can effectively improve the performance of the model.By combining these multiple schemes,we achieve the best results on the three data sets SDUMLA-HMT,FV-USM,and MMCBNU?6000.3.We propose a finger vein recognition algorithm based on deep hashing.We first combine the deep supervised hashing method to construct a finger vein recognition framework.During the training phase,we constrain the output of the neural network to ± 1.In the inference phase,we obtain the binary feature by the output binary-like feature and the sign function.We then use the hamming distance to evaluate the similarity of finger vein images.Experiments show that the proposed method has apparent advantages in storage consumption and retrieval speed while ensuring high retrieval accuracy.