一种雨雾背景的DeRF-YOLOv3-X目标检测方法

A DeRF-YOLOv3-X Object Detection Method for Rainy and Foggy Background

提出了一种新型的目标检测方法DeRF-YOLOv3-X(Derain and Defog-YOLOv3-Xception),将Xception引入YOLOv3网络以提高雨雾天气条件下行人和车辆的目标检测准确性。对于雨雾背景,分别采用残差网络和负映射结合的深度细节网络DDN和基于注意力机制的多尺度网络GridDehazeNet进行去雨去雾处理;采用Xception替换YOLOv3中Darknet-53网络,同时将回归损失函数由IoU改进为DIoU,提高特征提取能力以及框定位准确率。在公开数据集ImageNet上进行主干网络的测试;在实际场景数据集上进行YOLOv3-X网络和DeRF-YOLOv3-X网络的测试。实验结果表明,提出的DeRF-YOLOv3-X目标检测网络在雨天背景下mAP值提高了5.92%,达到54.99%;在雾天背景下,mAP值也提高了4.22%,达到49.07%。

A novel network named DeRF-YOLOv3-X(Derain and Defog-YOLOv3-Xception)is proposed.Xception is added into YOLOv3 network to improve the detection accuracy of pedestrians as well as vehicles in rainy and foggy weather.On the first stage,the image of rain and fog background is preprocessed,the residual network and the depth detail network(DDN)combined with negative mapping are used to decrease the influence of the rain.Besides,the multi-scale network of GridDehazeNet based on attention mechanism is used to desalinate the fog;Xception is also used to replace the Darknet-53 network in YOLOv3 and the regression loss function is modified from IoU to DIoU to better evaluate the extraction ability and frame positioning accuracy.The test of backbone networks is mainly based on public datasets ImageNet,and the test of YOLOv3-X and DeRF-YOLOv3-X networks is based on real-world datasets.The experimental results show that the proposed DeRF-YOLOv3-X target detection network improves the map values by 5.92%to 54.99%in rainy days,and 4.22%to 49.07%under the fog background.