Palmprint orientation field recovery via attention-based generative adversarial network

Abstract

Orientation field is the key foundation of palmprint feature extraction and recognition. However, due to the presence of numerous wide creases, the palmprint orientation field can hardly be accurately estimated by previous methods, especially in the thenar region, which still faces huge challenges. To solve this problem, we formulate palmprint orientation field recovery as an inpainting task and propose a palmprint orientation field recovery model named attention-based generative adversarial network. The deep generative architecture provides a powerful representation and an attention module guides the network to adaptively focus on the inpainting region. To avoid manually marking orientation field, we design a quality evaluation module to iteratively obtain pseudo labels for model training and incorporate palmprint abundant prior knowledge as extra supervision information. Palmprint identification results on public THUPALMLAB palmprint database show that our proposed algorithm improves the rank-1 recognition rate from 91.7% to 99.3%, which significantly outperforms the state-of-the-arts. Besides, we also compare the sensitivity of our algorithm to different optimization methods and noise distributions. Robust performance provides a reliable evidence for law enforcement and biometric identification.