DPHNet: Dual-Path Hybrid Network for Blurry Face Image Super-Resolution
Tailai Qiu, Yubao Yan · IEEE Access · 2024
Face images captured in the real world are usually corrupted by motion blur, which inevitably degrades the performance of current face super-resolution methods. To remedy this, in this study, a dual-path hybrid network with a CNN branch and a Transformer branch is developed. Specifically, the Transformer branch captures global features while the CNN branch focuses on extracting local features. Then, features extracted from these two branches are aggregated through a convolutional block attention module to capture spatio-channel correlation. With global features, blur cues can be well captured to facilitate clean images to be reconstructed. Experiments on the CelebA and FFHQ datasets demonstrate that the proposed network outperforms existing methods with notable margins.