U-PISRNet: A Unet-Shape Palmprint Image Super-Resolution Network

Palmprint has gained significant attention in recent years due to its reliability and uniqueness for biometric recognition. However, most existing palmprint recognition methods focus only feature representation and matching under an assumption that palmprint images are high-quality, while practical palmprint images are usually captured by various cameras under diverse backgrounds, heavily reducing the quality of palmprint images. To address this, in this paper, we propose a Unet-shape palmprint image super-resolution network (U-PISRNet) by learning and recovering multi-scale palmprint-specific characteristics of palmprint images. First, we project the palmprint images into the high-dimensional shallow representation. Then, we employ the transformer-based Unet-shape Encoder-Decoder architecture with skip-connections to simultaneously learn multi-scale local and global semantic features of palmprint images. Lastly, we reconstruct the super-resolution palmprint images with clear palmprint-specific texture and edge characteristics via two convolutional layers with embedding a PixelShuffle. Experimental results on three public palmprint databases clearly show the effectiveness of the proposed palmprint image super-resolution network.