摘要
无人机航拍图像的目标检测需要大量计算资源,导致其在移动端上的应用受到限制,为此提出一种轻量化目标检测网络Faster S。首先,在颈部网络设计了由部分卷积和组卷积组成的轻薄迅颈模块,用以有效聚合不同尺度的信息;其次,在检测头部分设计了一种高效、轻量化的单解耦头输出结构(SDHEAD),用以进一步分离目标检测中分类和定位两个子任务的特征编码。在航拍车辆数据集CARPK上的实验结果表明:与特征金字塔网络(FPN)结构相比,采用迅颈模块提高了模型的检测精度,减少了模型的参数量;相比于其他检测头,SDHEAD具有更强的特征辨别能力;Faster S模型检测的平均精度(AP_(0.5))达到了70.9%,较Yolo-FastestV2提高了12.7%,较Fastestdet提高了3.5%,推理时间仅为29.56 ms,较Nanodet-m降低了76%,模型参数量仅为0.335 M。
Object detection in drone aerial images require intensive computational resources,which limits the application on mobile terminals.Therefore,a lightweight target detection network Faster S is proposed.Firstly,a light and thin neck module composed of partial convolution and group convolution is designed in the neck network to effectively aggregate information of different scales.Secondly,an efficient and lightweight single decoupled head output structure(SDHEAD)is designed in the detection head part,which can further separate the feature encoding of the two subtasks of classification and localization in object detection.The results of experimental on the aerial vehicle dataset CARPK show that compared with the feature pyramid network(FPN)structure,the fast neck module improves the detection accuracy of the model and reduces the parameters of the model.Compared with other detection heads,SDHEAD has stronger feature discrimination ability.The average accuracy(AP_(0.5))of Faster S model detection reaches 70.9%,which is 12.7%higher than YoloFastestV2,3.5%higher than Fastestdet,and the inference time is only 29.56 ms,which is 76%lower than Nanodet-m.The parameter size is only 0.335 M.
作者
高宏伟
王雨桐
GAO Hongwei;WANG Yutong(Shenyang Ligong University,Shenyang 110159,China)
出处
《沈阳理工大学学报》
CAS
2024年第2期1-6,共6页
Journal of Shenyang Ligong University
基金
辽宁省重点科技创新基地联合开放基金项目(2021-KF-12-05)。
关键词
目标检测
航拍图像
轻量化
单解耦头
迅颈模块
object detection
aerial image
lightweight
single decoupled head
fast neck module