Abstract: Aiming at the problems of high missed detection rate, low detection success rate and large model size in traditional UAV（unmanned aerial vehicle） aerial image target detection algorithms, a new target detection method based on depthwise separable multiplex networks structure, DSM-YOLO v5（depthwise separable multiplex YOLO v5） is proposed. By adding a small target detection head with a size of 160×160 to the YOLO v5 network structure and performing residual connections with the high-level network to improve the detection ability of small objects. At the same time, the depthwise separable convolution algorithm is introduced in the Conv module, replacing ordinary convolution with depthwise separable convolution, which can effectively reduce the number of network parameters and reduce the model size. The experimental results show that the mAP@0.5 of target detection based on DSM-YOLO v5 network structure is 36.8%, which is 3.6 percentage points higher than that of YOLO v5s. The parameter volume and model volume are reduced by 21.1% and 19.7% respectively compared with YOLO v5s, which can be effectively applied to UAV aerial image target detection tasks.

Key words: DSM-YOLO v5, target detection, depthwise separable, YOLO v5, unmanned aerial vehicle（UAV） image

摘要： 针对传统无人机航拍图像目标检测算法存在漏检率高、检测成功率低、模型体积大等问题，提出一种新的基于深度可分离多头网络结构的目标检测方法DSM-YOLO v5（depthwise separable multiplex YOLO v5）。通过在YOLO v5网络结构上增加一个尺寸为160×160的小目标检测头，并将其与高层网络进行残差连接，以提升小目标检测能力，同时在Conv模块中引入深度可分离卷积算法，将其中的普通卷积替换为深度可分离卷积，能有效减少网络的参数量，降低模型体积。实验结果表明，基于DSM-YOLO v5网络结构的目标检测的mAP@0.5为36.8%，较YOLO v5s提高3.6个百分点，参数量和模型体积较YOLO v5s整体下降21.1%和19.7%，能够有效地应用于无人机航拍图像目标检测任务。

关键词: DSM-YOLO v5, 目标检测, 深度可分离, YOLO v5, 无人机图像