基于改进YOLOv4-tiny的无人机航拍目标检测

UAV Aerial Target Detection Based on Improved YOLOv4-tiny

针对无人机航拍图像中存在目标尺寸小、数量多和背景复杂等问题,提出了一种基于改进YOLOv4-tiny的无人机航拍目标检测算法。该算法在原有网络的基础上扩大了检测尺度范围,提高对不同尺寸目标的匹配程度,并利用深层语义信息自下而上地与浅层语义信息进行融合以丰富小目标的特征信息。同时引入注意力机制模块,在主干网络后的每个尺度上进行感兴趣区域特征信息的二次筛选,过滤冗余特征信息,保留重要特征信息。在无人机航拍数据集上进行对比实验,实验结果表明,所提算法在满足实时性的基础上,平均精确率比原网络提高了5.09%,具有较好的综合性能。

Aiming at the problems of small size,large number of targets and complex background in UAV aerial images,a UAV aerial target detection algorithm based on improved YOLOv4-tiny is proposed.On the basis of the original network,the algorithm expands the scope of detection scale,improves the matching degree for targets of different sizes,and fuses deep semantic information with shallow semantic information from bottom to top to enrich the feature information of small targets.Meanwhile,the attention mechanism module is introduced to conduct secondary screening of the feature information of the region of interest at each scale behind the backbone network.So as to filter the redundant feature information,and retain the key feature information.Compared with that of the original network,the average accuracy of the proposed algorithm is improved by 5.09% on the basis of real-time performance,and the experimental results show that the proposed algorithm has good comprehensive performance.