Edge Computing with Green Super-Resolution

Abstract

Image Super-Resolution (SR) aims to recover high-fidelity details from low-resolution inputs. It is a task inherently challenging due to its ill-posed, one-to-many mapping nature. Although deep learning (DL) approaches have achieved state-of-the-art performance, they often incur high computational overhead and lack mathematical transparency, hindering their deployment in resource-constrained edge environments. To address these limitations, we propose a novel green AI framework, called green super-resolution (GSR). GSR synergizes interpretability with efficiency. First, we introduce an easy-hard sample-space routing strategy. This conditional computation mechanism optimizes resource allocation by bypassing flat, easy regions and directing computational focus toward hard samples. Next, to address pixels in textured or edge areas, we propose an edge-assisted patch-extraction and alignment mechanism that yields consistent textural patterns in an unsupervised manner. Afterward, we adopt a feature ranking process to zoom into a discriminant feature subspace and generate more powerful features via the least-squares normal transform (LNT). Finally, the residuals are predicted using an efficient XGBoost regressor. Extensive experiments demonstrate that the proposed GSR method achieves highly competitive restoration quality while maintaining an exceptionally compact model size and significantly lower inference complexity compared to earlier SR benchmarks.