Adaptive Illumination Recovery for BackLit Images: Leveraging URetinex-Net and Iterative Prompt Learning

This paper presents a low-light image enhancement model built on image–text contrastive pre-training. The approach integrates an adaptive prior–based enhancement module with the CLIP semantic guidance framework to address the limitations of existing methods [19, 30]. Earlier low-light enhancement systems can align visual features with text supervision to some degree, yet they often struggle in scenes with severe darkness or strong backlighting. Other models improve image brightness effectively but do not fully utilize the broad and diverse priors available in CLIP for distinguishing illumination conditions [35]. Our method introduces a learnable illumination enhancement network that draws on the Retinex decomposition principle and incorporates adaptive physical priors. This design enables a more reliable separation of reflectance and illumination, allowing the network to handle challenging lighting variations. At the framework level, we adopt a CLIP-based prompt learning strategy and incorporate an iterative contrastive mechanism [3]. This allows the system to generate task-relevant guidance automatically, avoiding hand-crafted prompts or manual tuning. Experiments on the BAID low-light benchmark show that the proposed model achieves consistent improvements over prior work. Compared with CLIP-LIT, one of the strongest recent methods, our approach improves Peak Signal-to-Noise Ratio (PSNR) by 12.3% and Structural Similarity Index (SSIM) by 14.6%. These results demonstrate the effectiveness of our model in realistic low-light and backlit environments.