Low-Light Image Enhancement

Low-light image enhancement studies how to recover brightness, color, and structure from images whose weak signals are corrupted by sensor noise, blur, and uneven illumination. Methods commonly use Retinex-style illumination and reflectance decomposition, learned curves, color-space or frequency decoupling, and CNNs, Transformers, state-space models, diffusion, or flow-based restoration guided by physical priors. Training ranges from paired supervision on benchmarks such as LOL, SID, and LSRW to unpaired or zero-reference learning; events, RAW data, bursts, and auxiliary depth or semantic cues can provide additional information, while some systems optimize for detection or segmentation rather than appearance alone. Key challenges are generalizing from synthetic or paired data to diverse real scenes, balancing naturalness and color fidelity against pixel metrics, and delivering efficient, controllable enhancement without amplifying noise or erasing detail.