DSMISR: Differential Siamese Multi-scale Attention Network for Iris Image Super Resolution

Jin Hao, Shijie Lian, Suqi Li, Hua Li

Abstract: Iris images sampled in relaxed acquisition conditions usually have poor high frequency details, which severely affects the accuracy of iris recognition. Super resolution, as a kind of prior knowledge, can enhance the texture details for sequent recognition, but there are few relevant studies. In this work, we propose a learning-based super resolution neural network, i.e., Differential Siameses Multi-scale attention network for Iris super resolution (DSMISR). We develop a unique differential siamese multi-scale attention (DSMA) block, which extracts features via siamese structure but segments them into different scales. In this way, the module can generate multi-scale information in a consistent manifold, and utilize the differential information to improve the final performance. DSMA block simultaneously accelerates the information dissemination in the shifted window scheme and recovers more high frequency information. With this novel module, we reduced the computing cost by up to 25% compared to SwinIR. Extensive results show that DSMISR gets better PSNR and SSIM scores and decreases the ERR in iris recognition system effectively.