Real-time multi-domain collaborative low-light image enhancement for driving safety

Visual information quality in low-light environments is critical for ensuring autonomous driving safety. Under conditions such as nighttime driving, captured images suffer severe degradation, characterized by color distortion, noise interference, and loss of detail. This not only challenges the performance of visual perception systems but also limits the accuracy of subsequent advanced tasks like object detection, ultimately posing a potential risk to driving safety. To address this challenge, we propose a novel, efficient multi-domain collaborative low-light enhancement network (MCLE-Net) to meet the demands of real-time processing and high-fidelity restoration during low-light driving. The key component is the multi-domain attention block (MDAB), which collaboratively and explicitly processes spatial, channel, and frequency domain image features within a unified architecture. By integrating the fast fourier transform (FFT) into its attention mechanism, it precisely recovers high-frequency details critical for perception tasks while optimizing global contrast. Simultaneously, sparse spatial attention (SSA) and sparse channel attention (SCA) mechanisms adaptively focus on salient features with minimal computational overhead. Furthermore, the network incorporates a U-Net-inspired multi-scale feature extraction block (MSFEB) for progressive feature refinement. A gated fusion block (GFB) dynamically aggregates multi-scale information, improving feature utilization and representational capacity. Extensive experiments on multiple benchmark datasets demonstrate that MCLE-Net achieves an competitive balance between enhancement quality and computational efficiency by significantly improving feature discernibility and detail recovery, providing an efficient and reliable visual enhancement solution for ensuring low-light driving safety.