LAFE: A lightweight edge-aware nonlinear fusion framework for nighttime HDR image enhancement

Abstract

Enhancing nighttime images with high dynamic range (HDR) characteristics remains a formidable challenge due to the coexistence of extremely dark backgrounds and intense artificial light sources. Current deep learning models mostly target low-light image enhancement issues rather than nighttime HDR scenes, and their enhancement effects are highly dependent on the training set, rather than improving from a physical perspective. To address these issues, this paper proposes a lightweight Local Adaptive Fusion and Edge-aware nonlinear enhancement strategy (LAFE).The proposed framework integrates Retinex-based decomposition with a multi-scale fusion architecture. The core innovation lies in a gradient-driven power-law modulation function, which adaptively regulates fusion weights based on local structural saliency, effectively mitigating the ringing artifacts and over-exposure common in linear weighting schemes. Furthermore, a no-reference quality-driven optimization mechanism is introduced, utilizing the Natural Image Quality Evaluator (NIQE) as an objective function to automatically search for optimal modulation parameters via a proxy-image Pattern Search algorithm coupled with a dynamic early stopping strategy. Extensive experiments on five datasets have shown that LAFE can achieve satisfactory performance in both subjective perception and objective indicators. Quantitative evaluations demonstrate that LAFE outperforms the baseline fusion method, reducing the NIQE score from 3.65 to 3.38 and the BRISQUE score from 25.33 to 24.12. Compared to typical data-driven approaches, LAFE offers practical advantages as a lightweight, training-free method that is entirely independent of training data and GPU acceleration, achieving highly competitive performance without hardware overheads.