From Pixels to Perception: Generative and Adaptive Methodologies for Degraded Visual Scenarios

Enhancing images and videos in degraded visual scenarios is a vital task in computer vision, with applications across diverse fields such as medical imaging, robotics, and photography. Existing works have focused heavily on improving pixel-level quality in the final enhanced images through techniques like denoising, deblurring, and deraining. However, increasing evidence shows that while these pixel-level enhancement methods produce visually appealing results for human sense, they may not be optimal for downstream-task perception, such as object detection and semantic segmentation, which are essential for applications like robotics and autonomous driving. In this thesis, we aim to explore image and video enhancement from pixel-level enhancement for human sense to perception-level enhancement tailored for downstream computer vision tasks This will be achieved by developing innovative enhancement systems leveraging cutting-edge methodologies, such as generative models and domain adaptation. Specifically, this thesis explores four representative tasks: waterdrop removal for driving scenes in rainy scenarios, low-light image enhancement with generative models, enhancing HDR imaging with joint denoising and deblurring, adapting large multi-modal models to see and read in the dark. With extensive experiments, it gradually highlights the distinction between pixel-level enhancement for human sense and perception-level enhancement for downstream computer vision tasks, providing a clearer direction for future research on enhancement objectives.