Convolutional neural network based on multi-directional local coding for finger vein recognition

Convolutional neural networks (CNNs) have been widely used in biometrics due to their powerful image characterization capabilities. However, for small sample datasets like finger veins, CNNs with too deep or too wide structures are prone to overfitting problems. Moreover, complex networks require long training time and high-performance equipment, while simple networks have insufficient ability to express features. To address this problem, we present a lightweight model (Lmodel) and a feature integration model (FIModel), where the Lmodel can reduce the network parameters to save learning time, and the FIModel can ensure the model’s accuracy. More specifically, for LModel, we propose a new multi-directional local feature coding (MDLFC) to efficiently extract shallow features from the images, which will be input into the CNN to refine the features further. For the FIModel, we input the original image to the CNN for feature extraction and integrate it with the features extracted by LModel to enhance the feature representation. Surprisingly, compared with traditional CNN, even if FIModel incorporates Lmodel, it is still a lightweight network, which is also suitable for small sample datasets such as finger veins. The proposed system is validated on two publicly available databases and obtains superior performance over other state-of-the-art results.