CDBGrad-BlindSR: Collaborative Dual-Branch Network via Gradient Guidance for Efficient Blind Super Resolution

Abstract

Existing degradation kernel-based image super-resolution (SR) algorithms have achieved favorable performance in blind SR measurement. This addresses the limitation where bicubic kernel-based SR methods suffer performance degradation when the input image’s degradation kernel deviates from the assumed kernel. However, predicting the degradation kernel demands additional computational resources beyond the SR model. Moreover, imprecise degradation kernel estimation often results in restored images with visible artifacts and distorted structural details. To mitigate the above limitations, we introduce an efficient collaborative dual-branch network via gradient guidance for blind SR measurement. Concretely, without relying on degradation kernel estimation, we utilize the gradient spatial feature transform (GSFT) layer to enable mutual collaboration between features extracted from the restored branch and gradient prior features. These gradient prior features effectively capture the structural details of the input low-resolution (LR) image. To reduce model complexity, we adopt an information distillation mechanism in both channel and spatial attention mechanisms as the feature extractor, allowing the network to focus on essential features while bypassing redundant ones. Furthermore, to thoroughly exploit degradation priors during training without kernel estimation, we introduce a calibration mechanism. In this mechanism, the studied LR image degraded from the SR image by using the degradation kernel prior, is aligned with the ground-truth LR image under the constraints of $L1$ loss and a second-order gradient loss. Under this constraint, the SR image can be indirectly aligned with the high-resolution (HR) image. Extensive experimental results demonstrate that our network significantly economizes model parameters and computational costs by eliminating the degradation kernel estimation process. Meanwhile, it maintains competitive blind SR performance compared to other state-of-the-art (SOTA) methods.