基于ASPP-YOLOv4多尺度融合无人机图像目标检测  

UAV IMAGE TARGET DETECTION BASED ON ASPP-YOLOV4 MULTI-SCALE FUSION

作　　者：王玲 韩卓育 王鹏[1] 白燕娥[1] Wang Ling;Han Zhuoyu;Wang Peng;Bai Yan’e(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机应用与软件》2025年第2期190-195,235,共7页Computer Applications and Software

基　　金：中央引导地方科技发展基金吉林省基础研究专项(202002038JC)。

摘　　要：针对无人机视频图像背景复杂、小目标数量多、漏检错检率高的问题,提出一种基于改进YOLOv4的小目标检测算法。加入改进的注意力机制来加强关注小目标信息的能力;增加一个检测头并与主干网络的特征图进行融合来获取小目标的语义信息;使用改进的ASPP网络代替普通卷积块进行下采样以增大感受野,减少信息丢失。在VisDrone2019数据集上的实验结果表明,ASPP-YOLOv4的mAP较YOLOv4提升3.82百分点,显著地提升了小目标的检测精度。Aimed at the problems of complex background,large number of small targets and high missed and false detection rate of UAV video image,a small target detection algorithm based on improved YOLOv4 is proposed.An improved attention mechanism was added to enhance the ability to focus on small target information.A detection head was added and fused with the feature map of the backbone network to obtain the semantic information of small targets.The improved ASPP network was used to replace the ordinary convolution block for down sampling to increase the receptive field and reduce the loss of information.The experimental results on the VisDrone2019 dataset show that the map of ASPP-YOLOv4 is 3.82 percentage points higher than that of YOLOv4,which significantly improves the detection accuracy of small targets.

关 键 词：无人机视频图像 小目标检测 YOLOv4 多尺度融合 ASPP 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]