A Hyper-Dense Connection Net for finger multimodal recognition
Ao Li, Hui Ma, Kang Cao, Liangjun Xu · 2024
Recently, multimodal biometric recognition has garnered significant attention due to its superior accuracy and security. However, traditional deep neural networks face feature loss during information propagation, resulting in degraded network performance and training difficulties. To address this issue, we designed a novel Hyper-Dense Connection Net (HDCNet) for dual-modal biometric recognition tasks involving fingerprints and finger veins. HDCNet employs an innovative Hyper-Dense connection pattern that maximizes feature reuse, facilitating the efficient propagation of gradients and feature information. This connection pattern ensures that each network layer has access to a rich set of forward features, enhancing feature reuse capabilities. Concurrently, we designed a Deep Feature Extraction Block (DFE-Block) that leverages convolution operations to extract detailed image features within a local receptive field. HDCNet is capable of capturing complex biometric representations, significantly improving classification performance. Experimental results demonstrate that HDCNet achieves a recognition accuracy of 99.6% on the SDUMLA dataset. On the FVC-HKPU dataset, it achieves a recognition accuracy of 99.9%. These findings confirm the superior performance of our approach compared to other SOTA convolutional neural network models.