Efficient dual-attention fusion with gated feed-forward learning for image restoration under compound degradations

Image restoration is a fundamental task in computer vision, widely applied to low-light enhancement, image denoising and compression artifact reduction. Transformer-based restoration methods are effective at capturing long-range dependencies and modeling global features, but they often suffer from insufficient interaction between shallow and deep features, as well as high computational overhead. To address these limitations, this paper proposes an image restoration method named DAGF-Swin, based on dual-attention fusion and gated feed-forward collaboration. First, a dual-attention fusion (DAF) module is introduced to jointly leverage channel attention and spatial attention, enhancing the collaborative representation of shallow texture details and deep semantic information. Second, a gated convolutional feed-forward network (GConvFF) is designed to improve nonlinear modeling capability and feature selection efficiency within the feed-forward stage. Third, a global shifted-window mask caching mechanism (MaskCache) is developed to reduce redundant computations and lower overall computational cost. Experimental results on multiple test datasets demonstrate that the proposed method achieves superior performance, with notable improvements in the full-reference metrics PSNR, SSIM and DISTS, as well as the no-reference metric NIQE. Furthermore, compared with previous classical methods, our approach achieves better running efficiency. These results indicate that effective feature fusion and lightweight computation can improve practical image restoration performance.