Local and Global Feature Interaction Network for Partial Finger Vein Recognition

Abstract

Small imaging windows capture partial finger vein images, and these images have fewer vein vessels than full finger vein images. Most of the state-of-the-art recognition methods dealt with full images, which may not extract adequate features from partial images and therefore fail to achieve a satisfactory performance. This paper proposes a local and global feature interaction network for partial finger vein recognition. The proposed network has three dynamic feature interaction stages, and in each stage there are three modules, i.e., the local feature extraction module, the global feature extraction module, and the feature interaction module. In detail, the local module extracts multi-scale local features and makes a spatial feature fusion of the multi-scale features, the global module extracts the global context features, and the interaction module enhances the representation ability of each kinds of feature by a dynamic weight calculation. After three interaction stages, a spatial feature fusion is performed on the interacted local and global features. We built two partial databases based on two open full finger vein databases, and the experimental results on two partial databases show the effectiveness of our network.