改进YOLOv5s的轻量化无人机小目标检测

Improved YOLOv5s for Lightweight Unmanned Aerial Vehicle Small Target Detection

针对无人机航拍图像小目标检测时由于小目标像素值少,特征不丰富、难提取和易受环境干扰等局限性,容易导致模型漏检、精度低以及网络参数量过大等问题,提出一种改进YOLOv5s网络的无人机小目标检测模型,改进了YOLOv5s网络结构,在不增加小目标检测头的条件下减化网络结构,减少参数量,使模型更轻量化;提出了金字塔池化模块(ASPPF),将空洞卷积加入到YOLOv5s网络的SPPF模块,加强特征信息的空间不变性,提高了网络对小目标的感知能力;采用跨层上采样(CLAU)注意力模块,在上采样过程后将低分辨率深度特征与高分辨率浅层特征融合,提高了对小目标图像的检测效率;使用EIoU损失函数替换原CIoU损失函数,提升训练的收敛速度。在VisDrone2019数据集上的验证表明,改进后模型的mAP@0.5和mAP@0.50∶0.95分别为41.2%和23.4%,较原模型分别提升了7.2和4.7个百分点,参数量仅为原先的49%。

For small target detection in aerial images of UAVs due to limitations such as low pixel values lack of rich features difficulty in feature extraction and susceptibility to environmental interference it is easy to lead to missed detections low accuracy and excessive network parameter quantities.To solve the problems a small target detection model for UAVs based on improved YOLOv5s network is proposed.The YOLOv5s network structure is improved by reducing the network structure and parameter quantity without adding small object detection heads making the model lighter.A pyramid pooling module ASPPF is proposed which adds dilated convolution to the SPPF module of the YOLOv5s network to enhance the spatial invariance of feature information and enhance the spatial invariance of feature information.The perception ability of the network towards small targets is improved by adopting a Cross Layer Upsampling(CLAU)attention module.After the upsampling process the low-resolution deep features are fused with the high-resolution shallow features to improve the detection efficiency of small target images.The EIoU loss function is used to replace the original CIoU loss function to improve the convergence speed of training.Validation on the VisDrone2019 dataset shows that:1)The improved model performs well in mAP@0.5 and mAP@0.5∶0.95 with values of 41.2%and 23.4%respectively which are 7.2 and 4.7 percentage points higher than that of the original model;and 2)The number of parameters is only 49%of the original model.