基于NDM-YOLOv8的无人机图像小目标检测

Small Target Detection in UAV Images Based on NDM-YOLOv8

针对无人机航拍图像中小目标实例多、目标之间存在遮挡的现象,容易造成漏检、误检等问题,提出一种新的基于非跨步动态多头结构的小目标检测算法(Non-strided Dynamic Multihead YOLOv8,NDM-YOLOv8)。首先,设计了SConv模块,融合了非跨步卷积,尽可能地保留输入数据的判别特征信息,以降低小目标特征的细粒度信息的丢失;其次,设计了C2f-LSK模块,通过采用选择机制对空间特征进行有效加权,动态地调整感受野,灵活地捕捉不同尺度的特征和上下文信息,提高模型对小目标的关注度;最后,设计了P2小目标检测头,并和高层网络进行残差连接,减少小目标特征丢失,以强化算法对小目标特征的提取能力。实验表明,NDM-YOLOv8有效提高了对无人机图像中小目标检测精度。在公开数据集VisDrone2019上,NDM-YOLOv8比YOLOv8n在mAP_(0.5)提高了5.3百分点,mAP_(0.5:0.95)上提高了3.3百分点,对比其他模型,也取得了较优的检测效果,能够有效地完成无人机航拍图像中小目标检测任务。

In view of the phenomenon that there are many instances of small targets in UAV aerial images and there are occlusions between targets,which can easily cause problems such as missed detection and false detection,a new small target detection algorithm based on a non-strided dynamic multihead structure,NDM-YOLOv8(Non-strided Dynamic Multihead YOLOv8),is proposed.Firstly,the SConv module was designed,which integrated non-strided convolution and preserved the discriminative feature information of the input data as much as possible to reduce the loss of fine-grained information of small target features.Secondly,the C2f-LSK module was designed to effectively weight spatial features through a selection mechanism,dynamically adjust receptive fields,flexibly capture features and contextual information of different scales,and improve the model’s attention to small targets.Finally,a P2 small target detection head was designed and residual connected to the high-level network to reduce the loss of small object features and enhance the algorithm’s ability to extract small target features.Experiments show that the NDM-YOLOv8 effectively improves the detection accuracy of small targets in drone images.On the public dataset VisDrone2019,the NDM-YOLOv8 improved by 5.3 percentage points on mAP_(0.5) and 3.3 percentage points on mAP _(0.5:0.95) compared with YOLOv8n.Compared with other models,it also achieved better detection performance and can more effectively complete the task of small target detection in UAV aerial images.