Dual Representation Aggregation Network for Blind Image Super-Resolution via Iterative Bi-level Optimization

Blind image super-resolution (BSR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) images with unknown degradations. Existing BSR methods mainly follow two pipelines: LR prior-based reconstruction and SR posterior-based reconstruction. However, LR prior-based methods are sensitive to degradation and lack correction, and SR posterior-based approaches face limitations in generalization. From the perspective of aggregating both LR priors and SR posteriors, we propose a multi-branch network, Dual Representation Aggregation Network (DRANet) that constructs degradation aware branch and deviation correction branch hierarchically. Moreover, the end-to-end case reveals that the interaction between two branches poses a significant challenge, leading to model collapse and poor performance. To address this issue, we propose Iterative Bi-level Optimization (IBO), which decouples the DRANet learning problem into three sub-problems for staged training. We first optimize the parameters of two branches separately. Once each branch converges to a satisfactory performance, we conduct joint training through an iterative approach. Experiments on several blind degradation settings show that our DRANet produces promising results compared to state-of-the-art BSR methods on various benchmarks. Ablation studies indicate the effectiveness of dual representation aggregation and the stability of IBO, compared to both single and end-to-end strategies. The code is available at:https://github.com/para999/IBO.