基于航空图像的目标检测算法Trans_YOLOv5

Target Detection Algorithm Trans_YOLOv5 Based on Aerial Image

与自然图像的检测算法相比较,航空图像的检测存在目标角度随机、目标尺度变化剧烈、小目标密集、图像背景复杂等问题。针对这一系列难题,提出适用于航空图像检测的Trans-YOLOv5算法。修改YOLOv5算法中数据预处理模块以及后处理方法,增加一个目标角度标签的处理,使其适用于目标角度随机的航空图像。针对后续出现的边界问题,引入CSL(Circular Smooth Label,圆形平滑标签)将标签角度回归问题转换为分类问题,提高角度标签检测的精度。针对航空图像小目标检测问题,将Swin Transformer集成于YOLOv5框架中,提升模型对小目标的检测效果,并配合注意力机制模块,提高全局表征能力,使网络模型更加关注于待检测的目标对象。在DOTAv2.0航空图像数据集上的实验结果验证了所提方法的有效性,检测结果达到60.98%mAP,与原YOLOv5算法检测结果相比提高10.85百分点,与官网公布的竞赛最佳结果相比提高2.01百分点。

Compared with the detection algorithm of natural images,there are problems such as random target angle,sharp change of target scale,dense small targets,and complex image background in aerial image target detection.Trans-YOLOv5 algorithm suitable for aerial image detection is proposed to solve this series of problems.Modifying the data preprocessing module and post-processing method in the YOLOv5 algorithm to add the processing of a target angle label to make it suitable for aerial images with random target angles.CSL(Circular Smooth Label)is introduced to transform the label angle regression issue into a classification issue about the problem of boundary problems.Regarding the issue of small target detection in aerial images,we integrate Swin Transformer into the YOLOv5 framework to capture global semantic information,which improve the detection effect of the model on small targets,and cooperate with the attention mechanism module to improve the global representation ability,so that the network model pays more attention to the target object to be detected.The experimental results on the DOTAv2.0 dataset validate the effectiveness of the proposed method.The detection results reach 60.98%mAP,which is 10.85 percentage points higher than that of the original YOLOv5 algorithm and 2.01 percentage points higher than the competition results published on the official website.