Objective quality evaluation and enhancement of images affected by adverse weather conditions

Recently, there has been a surge in algorithms aimed at enhancing the perceptual quality of images degraded by adverse weather conditions such as snow, haze, and rain. However, limited attention has been given to evaluating the visual quality of these enhanced images or leveraging perceptual quality assessment metrics as loss functions to generate visually compelling results. This paper introduces the Adverse Weather Affected Image (AWAI) dataset, consisting of 2,800 real-world weather-degraded images and their enhanced counterparts, annotated with subjective perceptual quality scores. We also propose a novel no-reference weather-aware image quality assessment (NRWIQA) algorithm designed for such images to enable quantitative evaluation of enhancement quality. Unlike existing methods that use generic features trained on natural images, our approach extracts deep features sensitive to weather-induced artifacts. A multiscale feature extraction strategy, fused via a cross-attention mechanism, is used to capture degradations across different scales. These features are further refined using a transformer to predict perceptual quality scores. The experimental results demonstrate that the proposed NRWIQA algorithm significantly outperforms existing methods, achieving up to 17.61% improvement in SROCC, highlighting its effectiveness in assessing the quality of images affected by adverse weather conditions.