FreqAce-Net: Frequency-Aware Curve Estimation Network for Low-Light Image Enhancement in Autonomous Driving

Reliable visual perception systems, which are fundamental to autonomous driving, depend on high-quality images under challenging low-light environment. Existing enhancement methods often exhibit limitations in balancing noise suppression and detail preservation, leading to perceptual degradation. To address these issues, we propose FreqAce-Net, a low-light image enhancement network based on the Zero-Reference Deep Curve Estimation framework. We design a frequency-aware dynamic convolution module (FADC) to adaptively distinguish and process low-frequency noise and high-frequency details, and integrate this module into a progressive fusion network to guide more refined pixellevel curve parameter prediction. Moreover, we introduce a frequency reconstruction loss to guide the restoration of natural image statistics. The experimental results show that our method surpasses various state-of-the-art methods in image enhancement quality and performs excellently in the downstream 3D object detection task, indicating its potential for safe autonomous driving systems.