Efficient image denoising via green AI

Abstract

Image denoising aims to restore clean, high-fidelity visual structures from noisy inputs. Although deep learning (DL) approaches have dominated state-of-the-art benchmarks, they often incur heavy computational overhead, massive parameter counts, and a lack of mathematical transparency. To tackle these limitations, we propose a novel green AI framework for lightweight image denoising. Central to our framework is a hierarchical routing strategy that dynamically allocates computations based on local patch characteristics. First, input patches are parsed into smooth, texture, and edge regions using local gradient histograms. Smooth regions bypass processing to reduce computation, whereas texture and edge regions are further routed into easy, medium, and hard subcategories using lightweight classifiers based on residual magnitudes. Subsequently, samples in the texture and edge regions enter the Green Learning pipeline consisting of three sequential steps: unsupervised feature extraction, supervised feature selection and generation, and supervised decision learning for final residual prediction. Extensive experiments demonstrate that the proposed method achieves highly competitive denoising performance and visual quality while maintaining significantly lower computational complexity than existing benchmarks.