Synthetic generation of finger-vein region by feature fusion-based enhanced U-transformer for finger-vein recognition

Non-contact finger-vein recognition device offers high user convenience and minimizes hygienic issues, but it lacks a separate guide to support the finger region. Therefore, recognition performance can decline if recognized image includes areas not present in enrolled image. Previous approaches to address this issue still fail to overcome performance degradation when critical features are located in missing areas of the recognized image compared to the enrolled image. Therefore, this study proposes the method of synthetic generation of finger-vein region by feature fusion-based Enhanced U-transformer for finger-vein recognition. Enhanced U-transformer enhances recognition performance by outpainting missing finger-vein regions in recognized image using feature fusion-based U-shaped transformer. This improvement is achieved through modified cross-attention, residual layers, structural similarity index measure loss, and absolute positional embedding. The experiment utilized the Hong Kong Polytechnic University finger-image database version 1 and the Shandong University machine learning and applications-homologous multi-modal traits (SDUMLA-HMT) finger-vein database. Finger-vein recognition using Enhanced U-transformer achieved equal error rates of 3.01 % and 4.33 % in these databases, respectively, surpassing the performance of state-of-the-art methods. In addition, our Enhanced U-transformer demonstrates its ability to operate on embedded system with low computational resources.