Hierarchical Iris Image Super Resolution based on Wavelet Transform

Abstract

Iris images under the surveillance scenario are often low-quality, which makes the iris recognition challenging. Recently, deep learning-based methods are adopted to enhance the quality of iris images and achieve remarkable performance. However, these methods ignore the characteristics of the iris texture, which is important for iris recognition. In order to restore richer texture details, we propose a super-resolution network based on Wavelet with Transformer and Residual Attention Network (WTRAN). Specifically, we treat the low-resolution images as the low-frequency wavelet coefficients after wavelet decomposition and predict the corresponding high-frequency wavelet coefficients sequence. In order to extract detailed local features, we adopt both channel and spatial attention, and propose a Residual Dense Attention Block (RDAB). Furthermore, we propose a Convolutional Transformer Attention Module (CTAM) to integrate transformer and CNN to extract both the global topology and local texture details. In addition to constraining the quality of image generation, effective identity preserving constraints are also used to ensure the consistency of the super-resolution images in the high-level semantic space. Extensive experiments show that the proposed method has achieved competitive iris image super resolution results compared with the most advanced super-resolution method.