EAIM-Net: Environment-Aware Image Enhancement Network under Compound Degradations

Abstract

Outdoor images often suffer from compound degradations due to adverse weather conditions, atmospheric effects, and time-of-day illumination variations. Current restoration techniques focus on single degradation or require a specific degradation label or prompts at inference which restrict autonomous deployment. We introduce EAIM-Net, an Environment-Aware Adaptive Image Enhancement Network. It infers environmental factors from a degraded image, using a lightweight Environment Parameter Estimation (EPE) module, and then applies five specialist restoration filters. For training and evaluation, we propose the Weather-Text-Time (WTT) dataset, which contains 11,000 physics based paired samples in five weather conditions, four times of day and three illumination levels across 18 physically realised environmental states, for scenes containing text. Extensive experiments on the proposed WTT dataset and six public benchmarks demonstrate that EAIM-Net achieves competitive PSNR and SSIM on single-degradation benchmarks and significantly outperforms recent unified restoration methods under compound degradations. It includes low-light, rain, haze and glare, together with statistically significant gains (Wilcoxon signed-rank test, p < 0.001) over recent unified baselines and a favourable computational profile (9.7M parameters, 41.4 ms inference at 256×256) suitable for real-time deployment. The code is available at www.github.com/Arni-123/EAIM-Net