Lightweight parallel feedback network based on CRL with policy transfer and enhancement for image super-resolution

Abstract

Image super-resolution (SR) is essential in applications such as surveillance, medical imaging, and remote sensing, but existing deep learning (DL) models often require high computational resources and struggle to recover fine details in lightweight architectures. Although feedback and attention based methods have shown improvements, they still lack an effective combination of efficient feature refinement, edge enhancement, and low parameter complexity. To address this gap, we propose a lightweight parallel feedback network (LPFN) that combines three key components: a feedback block for repeated feature refinement, a dispersion-aware attention residual block (DARB) for highlighting important spatial and channel details, and EdgeNet for edge sharpening for sharper boundaries. These components are supported by curriculum reinforcement learning (CRL), an adaptive training strategy that gradually improves the model’s learning behavior. Instead of relying on a fixed loss function, LPFN uses a dynamically learned global feedback loss to refine reconstruction quality at each stage. Experiments on DIV2K and Flickr2K show that LPFN achieves higher PSNR and SSIMscores while keeping the model lightweight and efficient. This study emphasizes an effective lightweight feedback framework, an enhanced attention and edge-refinement mechanism, and an adaptive learning strategy that improves both accuracy and stability under different degradation conditions.