Enhancing Hand Feature Recognition with Limited Data via Multi-Scale Reconstruction Attention

Compared with Single Modal biometric systems, Multimodal systems can improve recognition accuracy and security. However, entirely using shallow detail features and improving the recognition rate when data samples are scarce remains a challenge in multimodal recognition. To this end, we propose the RMS-CNN model, a multi-scale reconstruction convolutional neural network trained on small-sample hand Multimodal data. The model uses a multi-scale attention mechanism to capture shallow local features effectively; simultaneously, an adaptive fusion method is used to balance the fusion of features from each modality. Experiments show that RMS-CNN can maintain high stability, generalization, and accuracy even with small training data.