Spectral-and Contrast-Aware Low-Light Image Enhancement for Robust Autonomous Driving

In low-light conditions, autonomous driving perception systems suffer feature loss, which results in deteriorating image quality and leads to the failure of downstream tasks like object detection and semantic segmentation. However, due to the sensor spectral bias and spatial illumination variation, the previous deep learning-based low-light enhancement algorithms result in the problems such as color maladjustment and uneven contrast. To tackle these challenges, we propose a novel enhancement framework which integrates the spectral compensation for color correction (SC-CCM) and region-adaptive contrast enhancement (RA-CE) modules. The SC-CCM module estimates spectral attenuation and corrects for it in RGB channels, thus reducing color distortions introduced by non-uniform sensor responses. In contrast, the RA-CE method dynamically tunes the contrast locally through learned illumination segmentation, enabling a consistent enhancement across each region. The model is based on an adapted Transformer architecture, and our method achieves currently the new-record performance on standard benchmarks as well as robustness on real-world driving conditions for the perceptual quality.