Adaptive Dehazing YOLO for Object Detection

While CNN-based object detection methods operate smoothly in normal images, they produce poor detection results under adverse weather conditions due to image degradation. To address this issue, we propose a novel Adaptive Dehazing YOLO (DH-YOLO) framework to reduce the impact of weather information on the detection tasks. DH-YOLO is a multi-task learning paradigm that jointly optimizes object detection and image restoration tasks in an end-to-end fashion. In the image restoration module, the feature extraction network serves as an encoder, and a Feature Filtering Module (FFM) is used to remove redundant features. The FFM contains an Adaptive Dehazing Module for image recovery, whose parameters are quickly calculated using a lightweight Cascaded Partial Decoder. This allows the framework to make use of weather-invariant information in hazy images to extract haze-free features. By sharing three feature layers at different scales between the two subtasks, the performance of the object detection network is improved by the use of clear features. DH-YOLO is based on YOLOv4 and forms a unified, end-to-end model with the above modules. Experimental results show that our method outperforms many advanced detection methods on real-world foggy datasets, demonstrating its effectiveness in object detection under adverse weather conditions.