31 December 2025 in Neural Networks
29 January 2026 in IEEE Access
Images captured in low-light conditions often suffer from overexposure, amplified noise, and uneven illumination. Supervised enhancement methods tend to overfit specific scenes, while unsupervised ones struggle to model degradations without reference images. To address these challenges, we propose a zero-reference Retinex framework. This framework integrates bilateral grid learning with an illumination-guided cross-attention mechanism, enabling stronger multi-scale feature extraction and enhanced illumination perception. We construct a three-dimensional spatial–luminance representation to model illumination distributions and introduce an illumination-guided cross-attention module to enhance feature alignment. Dynamic slicing and modulation further refine illumination, improving global consistency while preserving structural details. On the LOLv1 dataset, our method achieves 19.91 dB PSNR and reduces LPIPS by 14.34% compared with recent zero-reference approaches. Extensive experiments across multiple public datasets further confirm the strong PSNR, SSIM, and LPIPS performance of our method over state-of-the-art approaches.