Introduction
Intelligent Transportation Systems (ITS) are increasingly reliant on Ultra-High-Definition (UHD) 4K cameras to monitor traffic, detect accidents, and manage urban infrastructure. However, the effectiveness of these systems drops sharply at night. Low-light conditions introduce significant sensor noise and reduce contrast, obscuring critical details such as license plates and lane markings. Existing enhancement methods often struggle with the massive computational load of 4K video or fail to preserve the sharp edges necessary for downstream AI tasks like object detection.
The DDNet Framework
To address these challenges, the authors propose the Double Domain Guided Network (DDNet). DDNet is specifically designed to handle 4K surveillance footage in real-time by processing images in two parallel domains: color and gradient.
The architecture utilizes a sophisticated encoder-decoder structure to balance quality and speed:
- Double Domain Processing: The task is split between the Coarse Enhancement Module (CEM), which restores global brightness and color, and the Gradient Enhancement Module (GEM), which focuses on high-frequency structural details.
- LoG-based Guidance: To ensure the network enhances real edges rather than sensor noise, a Laplacian of Gaussian (LoG) operator is used. By smoothing the image before calculating gradients, the model becomes robust to the "salt and pepper" noise typical of low-light sensors.
- Self-calibrated Convolutions (ScCAM): To maintain real-time performance on 4K data, the model incorporates a lightweight attention mechanism that focuses on informative regions without the heavy overhead of standard Transformers.
Key Results and Impact
DDNet achieves a breakthrough in the "speed vs. quality" trade-off. It can process 4K (3840 × 2160) images at approximately 37 frames per second (FPS) on an Nvidia A40 GPU, making it viable for live surveillance feeds. This significantly outperforms many state-of-the-art models that require over a second per frame for UHD content.
Beyond visual quality, the enhancement significantly boosts the performance of downstream AI models. When using DDNet as a pre-processor, object detection accuracy (via YOLOv5) and scene segmentation (via DAFormer) improved substantially compared to raw footage or other state-of-the-art enhancement techniques. This proves that preserving high-frequency details in the gradient domain is essential for automated traffic analysis.
Conclusion
DDNet provides a robust solution for real-time UHD surveillance in challenging lighting. By prioritizing gradient information and employing efficient architectural components like ScCAM, it ensures that "smart" transportation systems remain functional around the clock. Future work will focus on reducing model size for edge deployment and addressing motion blur in high-speed traffic scenarios.
Summary updated