Joint face super-resolution and deblurring using multi-task feature fusion network
Y. Cui, Chao Tang, Qian Huang · IET conference proceedings. · 2023
Face image is an essential category of images. However, due to many factors, face images are often accompanied by one or more degradations, such as low-resolution and motion blurring. To solve these two problems simultaneously, we propose a new multi-task feature fusion network based on double branches, which can directly obtain high-resolution and clear face images from low-resolution and blurred face images. The accumulation of errors is effectively avoided by connecting the super-resolution feature extraction branch and the deblurring feature extraction branch in parallel. Adding the hybrid attention mechanism enhances feature selection at both channel and spatial levels. In addition, we propose a new multi-scale feature fusion module, which effectively integrates the features of the two tasks. In the loss part of the network, we also propose a local loss to guide the network to restore clearer facial details. Experiments show that our method can well increase image resolution and remove blurring. The restored image performs well in objective image evaluation indicators and advanced face visual tasks.