ECA-URetinex: a lightweight image enhancement algorithm for low-light autonomous driving

To address the significant challenges of image detail loss, severe noise interference, and computational efficiency for autonomous driving in extreme low-light conditions, this paper proposes ECA-URetinex, a highly efficient image enhancement model. The model is based on an interpretable Retinex Deep Unfolding framework, unfolding the optimization process of the reflectance R and illumination L components into iterative network stages, where the mechanism is integrated into the prior modules for their estimation. With a minimal parameter increase (only 0.08 M additional parameters), the integrated ECA module adaptively reinforces channel-wise feature. On the LOL dataset, the proposed ECA-URetinex model outperforms mainstream algorithms like URetinex-Net, KinD++, and URetinex-Net++ in terms of image quality. Specifically, it achieves the highest Peak Signal-to-Noise Ratio (PSNR) of 22.90 dB, the highest Structure Similarity Index Measure (SSIM) of 0.886, and the lowest Learned Perceptual Image Patch Similarity (LPIPS) of 0.112. Generalization experiments on autonomous driving scenes demonstrate that ECA-URetinex enhanced images significantly improve the accuracy of downstream object detection tasks, boosting the mAP@0.5 of the YOLOv8s model by 15.9%. This demonstrates the high robustness and efficiency advantages of the ECA-URetinex model, providing a highly practical preprocessing solution for all-weather autonomous perception systems.