改进YOLOv7的无人机图像小目标检测算法

Improved YOLOv7 Based UAV Image Small Target Detection Algorithm

针对现有的目标检测方法在识别无人机航拍图像中的小目标时精度较低且存在漏检或误检的情况,提出一种基于改进YOLOv7的小目标检测算法。首先,改进网络检测头,去掉原本用于检测大物体的20×20尺寸的检测头,增加了一个160×160尺寸的针对于小目标的检测头,并相应修改网络特征融合的路径;其次,重构SPPCSPC结构,裁剪卷积层并更改池化结构、以降低模块复杂度,加快网络收敛速度;然后,用内容感知特征重组(CARAFE)算子替换原上采样结构,减少图像在上采样时图像信息的丢失,最大化保留输入图像的局部和角落信息;最后,改进ELAN模块,轻量化主干网络的同时提高网络对小尺度目标的敏感度。在公开数据集VisDrone2019上进行实验,改进后模型的mAP50值达到56.6%,相较于原YOLOv7模型提升了2.9个百分点,且参数量减少了33%。

Aiming at the problems of low accuracy,missing detection and false detection in the identification of small targets in UAV aerial images by existing target detection methods,a small target detection method based on improved YOLOv7 is proposed.Firstly,the network detection head is improved,the 20×20 detection head originally used for detecting large objects is removed,a 160×160 detection head for small targets is added,and the network feature fusion path is modified accordingly.Secondly,the SPPCSPC structure is reconstructed,the convolutional layer is clipped and the pooling structure is changed to reduce module complexity and speed up network convergence.Then,the original upsampling structure is replaced by the Content-Aware ReAssembly of FEatures(CARAFE)operator to reduce the loss of image information during upsampling and maximize the preservation of local and corner information of the input image.Finally,the ELAN module is improved to lighten the backbone while improving the sensitivity of the network to small-scale targets.Experiments are carried out on the public dataset VisDrone2019,and the mAP50 of the improved model reached 56.6%,which is 2.9 percentage points higher than that of the original YOLOv7 model,and the parameters are reduced by 33%.