基于注意力机制的航拍图像目标检测算法  

Attention Mechanism-Based Object Detection Algorithm in Aerial Images

作　　者：白宗宝 张俊举[1] 高原 胡友成 Bai Zongbao;Zhang Junju;Gao Yuan;Hu Youcheng(School of Electronic and Optical Engineering,Nanjing University of Science&Technology,Nanjing 210094,Jiangsu,China;School of Electronic and Optical Engineering,Nanjing University of Science and Technology ZiJin College,Nanjing 210023,Jiangsu,China)

机构地区：[1]南京理工大学电子工程与光电技术学院,江苏南京210094 [2]南京理工大学紫金学院电子工程与光电技术学院,江苏南京210023

出　　处：《激光与光电子学进展》2023年第12期312-322,共11页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61971386)。

摘　　要：针对现有基于水平视角图像的目标检测网络在无人机航拍图像上误检率和漏检率高的问题,提出一种基于改进注意力机制的航拍图像目标检测算法。首先,在Faster R-CNN主干网络输出端引入一种三叉戟注意力机制,分别提取三路池化层和三路扩张卷积层的多模式信息及多尺度特征信息进行压缩,实现特征通道和空间像素区域的权重再分配。其次,针对航拍图像目标的分类和边界框回归,引入一种双头检测机制,充分利用目标的语义信息和空间位置信息。在相关数据集上对所提算法进行评测,并与其他目标检测算法进行对比。结果表明所提算法的平均精度均值得到显著提升,在不同场景下的无人机航拍图像目标检测中获得更好的效果。A target detection algorithm for aerial images based on an improved attention mechanism is suggested to address the issue that the existing object detection network based on horizontal view images has a high false-positive rate and a high miss rate in aerial images.First,a trident channel and spatial attention module that extracts multi-mode and multi-scale characteristic map data of three-branch pooling layers and three-branch dilated convolution layers is added at the output of the Faster R-CNN backbone network so as to compress the data,thereby redistributing the weight of feature channels and spatial pixel regions.Second,a double-head detection mechanism is employed for the classification of the objects and bounding box regression in the aerial image to fully utilize the semantic and spatial location information.The suggested algorithm is further assessed on relevant datasets and contrasted with other object detection algorithms.The results indicate a significant enhancement of the mean average precision of the suggest algorithm,leading to better target detection for unmanned aerial vehicle images in various scenes.

关 键 词：机器视觉 无人机 目标检测 注意力机制 双头检测机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]