UAVAI-YOLO:无人机航拍图像的小目标检测模型

UAVAI-YOLO:dense small target detection algorithm based on UAV aerial images

针对无人机航拍图像目标检测效果差的问题,提出改进的UAVAI-YOLO模型。首先,为使模型获得更加丰富的语义信息,使用改进可变形卷积网络(deformable convolutional networks,DCN)替换原骨干(back‐bone)网络部分通道到像素(channel-to-pixel,C2f)模块原始卷积。其次,为增加P2特征层而不增加模型参数量,提出Conv_C模块将骨干网络输出通道降维,同时避免通道降维导致的语义信息丢失,使用改进ODConv卷积替换颈部(neck)部分C2f模块原始卷积。然后,为充分利用上下文语义信息,引入双向特征金字塔网络(bidirectional feature pyramid network,BIFPN)。最后使用Wise-IoU替换原始损失函数,提高模型目标检测框的准确性。在公开的VisDrone2019数据集和UAVDT数据集的实验结果表明,UAVAI-YOLO模型相比于原YOLOv8n模型mAP@0.5分别提升了4.4%和1.1%。与其他主流目标检测模型相比具有较高的检测性精度。

An improved UAVAI-YOLO model was proposed to address the problem of poor target detection in UAV aerial images.Firstly,in order to obtain richer semantic information for the model,the original convolution of the C2f module of the original backbone part was replaced with the improved DCN convolution.Secondly,in order to increase the P2 feature layer without increasing the number of model parameters,the Conv_C module was proposed to downscale the output channel of the backbone network,and at the same time,because of avoiding the loss of semantic information due to channel downsizing,the original convolution of the C2f module in the neck part was replaced by the improved ODConv dynamic convolution.Then,the BIFPN module was introduced to make full use of the contextual semantic in‐formation.Finally,Wise-IoU was used to replace the original loss function to improve the accuracy of the model target de‐tection frame.Experimental results on the publicly available VisDrone2019 dataset and UAVDT dataset showed that the UAVAI-YOLO model improves 4.4%and 1.1%compared to the original YOLOv8n model mAP0.5,respectively,high de‐tectability accuracy compared to other mainstream object detection models.