TAFCNet: Task-aware feature coupling network for joint image motion deblurring and super-resolution

Abstract

Joint image motion deblurring and super-resolution aims to restore blurry low-resolution images with respect to image details and spatial content. Existing joint methods usually directly fuse the features but neglect the inherently different feature requirements and reconstruction objectives of the two tasks, resulting in failures of blurry texture reconstruction and inaccurate recovery of authentic details. To address this issue, we propose a joint degradation model grounded in the visible-light imaging process, which unifies motion blur and low sampling frequency degradation to characterize the linear relationship between features of deblurring and super-resolution. Guided by this model, we propose a novel Task-Aware Feature Coupling Network (TAFCNet) to enable restoration of local non-uniformly blurred regions and accurate reconstruction of structurally consistent content. Specifically, our method employs Stacked Nonlinear Feature Separator (SNFS) branch to enhance the representation of local information, and a High-Low Frequency Feature Mixer (HLFFM) branch operating in the frequency domain to obtain texture features from low-resolution images. To adaptively couple dual-task features, we further propose a Task-Aware Gating (TAG) module to mitigate feature conflicts. Extensive experiments demonstrate that our TAFCNet outperforms state-of-the-art methods in terms of objective metrics and subjective quality, achieving superior reconstruction performance and stronger cross-dataset generalization capabilities.