基于YOLO-v7的无人机航拍图像小目标检测改进算法

Enhanced Algorithm for Small Target Detection in UAV Aerial Images Based on YOLO-v7

在无人机航拍图像目标检测任务中传统目标检测算法实时性与精确性差,且原始YOLO算法对小目标的误检、漏检率较高。航拍图像对视角、图像数据量、目标尺度等方面的要求较高,与普通图像有显著差异。鉴于此,针对无人机航拍图像小目标检测难题,提出一种基于YOLO-v7的改进算法FCL-YOLO-v7。首先,添加小目标检测层,改进特征提取网络结构与先验框配置;其次,用FReLU激活函数替代原有的SiLU激活函数;再次,在骨干网络中添加CBAM注意力机制;最后,结合公开数据集与自主采集的无人机航拍图像构建小目标数据集。实验结果表明,改进算法在无人机航拍图像小目标数据集上的精确率比原始算法提高6.7%,比YOLO-v3提升7.3%;召回率比YOLO-v5高3.3%。

In the UAV aerial image target detection task,the traditional target detection algorithm is poor in real-time and accuracy.The orig-inal YOLO algorithm has a high error detection and omission rate for small targets.The requirements of aerial image are higher in view angle,image data amount,target scale and so on,which are significantly different from ordinary images.Therefore,an improved algorithm based on YOLO-v7,FCL-YOLO-v7,is proposed to solve the problem of small target detection in UAV aerial images.First,add small target detection layer,improve the feature extraction network structure and prior frame configuration;Secondly,the SiLU activation function is replaced by FReLU activation function.Thirdly,CBAM attention mechanism is added to the backbone network;Finally,the small target data set is con-structed by combining the open data set and the autonomous UAV aerial images.The experimental results show that the accuracy of the im-proved algorithm is 6.7%higher than that of the original algorithm and 7.3%higher than that of YOLO-v3.The recall rate is 3.3%higher than the YOLO-v5.