Prior-Informed Visual Enhancement and 3D Reconstruction under Challenging Illumination

Abstract

Low-light image enhancement (LLIE) and illumination-robust 3D scene reconstruction remain fundamental challenges in computer vision. Images captured under adverse illumination suffer from detail loss, amplified noise, and color shifts, degrading visual quality and hindering recognition, navigation, and scene understanding. Despite rapid progress, many LLIE pipelines still depend on image-to-image mappings trained on limited or synthetic datasets, and neglect auxiliary priors available from external data or physics. As a result, they overfit to seen degradations and break under heavy noise, motion, or extreme dynamic range. Meanwhile, modern 3D scene representation techniques, such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), rely on high-quality multi-view images and deteriorate severely with dark inputs. This thesis investigates how external priors (learned prior from normal-light images), perceptual knowledge extracted from vision–language models, and illumination-invariant physical cues, can be integrated to improve deep learning approaches in both 2D visual content enhancement and 3D scene reconstruction.

For visual content enhancement, two novel approaches, named GLARE and GPP-LLIE, are proposed to explore generative priors for low-light image enhancement. In \textbf{GLARE}, to mitigate the ambiguity of the enhancement process, the VQGAN with a discrete coodebook is employed to learn latent priors from normal-light images. Then, by aligning low-light features to this small proxy space via latent normalizing flow and adaptively transforming input features into the decoding process, GLARE restores missing structures and improves perceptual quality under extreme degradations. In \textbf{GPP-LLIE}, we first develop a pipeline to extract perceptual priors from large vision-language models. Then, these priors are integrated into the diffusion process as perceptual guidance, achieving results with high perceptual quality and demonstrating good generalization across real-world scenarios.

For 3D scene reconstruction under adverse illumination conditions, we present \textbf{LITA-GS}, an illumination-agnostic 3D Gaussian Splatting framework. By incorporating \textit{physical illumination-agnostic priors} with progressive denoising, LITA-GS achieves reference-free novel view synthesis that remains consistent and stable even under adverse lighting.

Together, these three contributions demonstrate the value of \textbf{codebook-based latent priors, perceptual quality priors, and physical illumination-agnostic priors} in addressing challenges across 2D and 3D vision. Extensive experiments confirm improvements in fidelity, perceptual realism, and robustness, establishing a foundation for future research on illumination-robust visual systems.