MCLL-Diff: Multiconditional Low-Light Image Enhancement Based on Diffusion Probabilistic Models

Abstract

Due to the inherent limitations of camera sensors in capturing adequate light under low-light conditions, images often suffer from various degradation issues, such as illumination imbalances, artifacts, and noise. While generative model-based methods have made remarkable progress in low-light image enhancement (LLIE), they still face challenges such as unstable training and inconsistent generation quality. To address these challenges, we introduce MCLL-Diff, a novel multiconditional LLIE method based on diffusion probabilistic model (DPM). MCLL-Diff retains the forward process of DPM but introduces a unique multiconditional noise predictor (MCNP) in the reverse process. We first propose a learnable operator module (LOM) to enrich the prior knowledge incorporated in the reverse process. Then, we use MCNP to effectively integrate prior knowledge, low-light images, intermediate variables, and time steps to accurately predict noise. To validate the effectiveness of MCLL-Diff in high-level computer vision tasks, we construct a large-scale nighttime vehicle model (NVM) dataset from real-world nighttime street scenarios. Extensive experiments on benchmark datasets demonstrate MCLL-Diff’s superiority in both generalization performance and visual quality. Specifically, we achieved a significant improvement of 0.1 dB in peak signal-to-noise ratio (PSNR) metric on the VE-LOL dataset, and a notable increase of 0.76% in Top-1 accuracy when applied to object recognition on the NVM dataset.