Contrast and Visibility Enhancement of Weather-Degraded Images Using Dual-Branch CNN and Transformer with Perceptual Loss for ADAS

Engineering notes

The model was tested on real-world driving image datasets such as BDD100K and KITTI Foggy Datasets and compared with state-of-the-art dehaze networks and general weather condition restoration networks.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Weather can be either poor or good. Poor weather, includING fog, hazE, rain, or low light, can cause dramatic degradation of image perception in road-level situations, leading to with significant performance loss in camera-based Advanced Driver-Assistance Systems (ADAS), Although traditional improvement techniques relying on Convolutional Networks (CNNs) cannot effectively preserve global context in image appearance improvement, techniques using transformers show high computational costs. This restricts their application as real-time system efficiency becomes critically important. In this paper, we propose a solution using the Dual-Branch CNN Transformer, which uniformly utilizes localized spatial features extraction together with global semantic modeling using parallel experience sharing of Convolutional Networks and Self Attention Mechanisms. An adaptive gated fusion module integrates these complementary local and global representations through learnable spatial weighting, while perceptual-loss-guided optimization emphasizes texture fidelity, structural consistency, and visual realism. The model was tested on real-world driving image datasets such as BDD100K and KITTI Foggy Datasets and compared with state-of-the-art dehaze networks and general weather condition restoration networks. The proposed model achieved a PSNR of 36.5 dB, an SSIM of 0.962, and an LPIPS of 0.081 while recording an inference latency of 42 ms/frame, corresponding to 23.8 FPS (~24 FPS) on an NVIDIA RTX 4090 GPU. Qualitative evaluation further demonstrated improved restoration of lane boundaries, vehicle contours, and overall scene coherence under adverse weather conditions. These findings indicate that the proposed framework provides an efficient and perceptually robust solution for visibility enhancement in autonomous driving scenarios.

Links and sources

Comments