Target Detection Algorithm for Foggy Scenes Based on Twin Network

In real scenes, the quality and contrast of images are reduced due to the uneven state of haze caused by factors such as humidity, dust, and aerosols in the air. In this scenario, it is difficult for a general detection network to detect targets in the image. Thus, a dual subnet based on multi-task collaborative training (DSMCT) is proposed. Firstly, during the training stage, GCANet is used as the supervisory network for YOLOX, which promotes the extraction of clean information by YOLOX in foggy scenes. During the testing stage, only the YOLOX branch is activated to ensure the detection speed of the model. Secondly, the deformable convolution module is used to improve the gated context aggregation network (GCANet), enhancing the model's ability to capture details of non-homogeneous fog. Finally, we introduce the Coordinate Attention mechanism into the Vision Transformer and redesign the backbone network of YOLOX to enhance the network's feature extraction capability for deep semantic information. The mAP of DSMCT reached 86.56% on the artificially hazy dataset (FOG_VOC), and it reached 62.39% on the real hazy dataset (RTTS). This is respectively 2.27% and 4.41% higher than the current state-of-the-art detection models. The experiments show that the improved DSMCT network has high practicality and effectiveness for object detection in real hazy scenes.