WAMT_DN: A Weather-Aware Multi-Task Detection Network for Object Recognition Under Adverse Weather Conditions

Object detection under adverse weather conditions remains a critical yet challenging task in computer vision, particularly for real-world applications such as autonomous driving and intelligent surveillance. The performance of conventional detection algorithms often degrades significantly in environments affected by rain, fog, or low light. To address this issue, this study proposes a novel weather-aware multi-task detection framework named WAMT_DN, to enhance detection accuracy. Based on the Faster Region-based Convolutional Neural Networks (R-CNN) architecture, the proposed method integrates weather recognition and domain classification tasks into a unified object detection pipeline. Specifically, two auxiliary branches are designed to predict weather types and domain labels. These branches are trained jointly with the leading detection network, enabling the model to learn more weather-invariant and generalizable features through multi-task supervision. Furthermore, the training process adopts mixed-precision optimization to accelerate computation and reduce memory consumption. Experimental results show that WAMT_DN achieves more stable convergence and lower detection losses than classical baselines such as Faster R-CNN and RetinaNet. Introducing auxiliary tasks significantly improves the performance in weather-degraded scenarios, confirming the effectiveness of multi-task learning in enhancing detection robustness.