VARSD-Net: Vision Adaptive Robust Scene Detection Network for Rainy and Low-Light Conditions

Reliable vehicle detection in complex traffic environments under rain and low-light is hindered by severe visual degradations that often lead to catastrophic feature collapse in real-world scenarios. To address this, we propose Vision Adaptive Robust Scene Detection Network (VARSD-Net), an end-to-end robust framework designed to bridge the gap between low-level feature refinement and high-level semantic recognition. Unlike traditional methods that often rely on computationally expensive pre-processing, VARSD-Net introduces a plug-and-play, training-only Multi-scale Refinement Network (MRNet) to stabilize latent information for degradation-invariant features without incurring inference-time latency. Building upon this refinement, we introduce the Adaptive Detection Network (ADNet), which integrates Linear Deformable Convolution (LDConv) and an Adaptive Attention Feature Aggregation (A2C2f-GCD) module. By utilizing Gated Cross-Domain Integration with channel-spatial attention, this module maximizes representational capacity and dynamically adjusts feature extraction pathways to preserve the details of obscured and distant targets. Furthermore, a weighted Bidirectional Feature Pyramid Network (BiFPN) facilitates cross-scale information exchange, effectively overcoming the resolution loss inherent in degraded inputs. Extensive evaluations on the RainCityscapes, RaidaR, ExDARK, and NOD datasets demonstrate that VARSD-Net significantly outperforms a diverse set of state-of-the-art detectors. With a parameter count of 8.9 M and a computational cost of 20.6 GFLOPs, our model sustains a competitive high-speed inference rate of 59 FPS. These results validate VARSD-Net as a robust expert system for ensuring the localization precision and reliability essential for safety–critical perception in real-world autonomous driving applications.