融合BotNet的遥感图像目标检测  

Target detection in remote sensing images by fusing BotNet

作　　者：赵精莹 郝晓丽[1] ZHAO Jing-ying;HAO Xiao-li(College of Computer Science and Technology,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学计算机科学与技术学院,山西晋中030600

出　　处：《计算机工程与设计》2024年第10期3026-3032,共7页Computer Engineering and Design

基　　金：国家自然科学基金面上基金项目(62072326)。

摘　　要：为解决遥感图像目标小、多尺度、目标背景复杂等问题,提出一种Bottleneck Transformer目标检测网络,在YOLOv5s模型上用“CNN+Transformer”架构代替最后一个残差块中的C3卷积操作,利用空洞卷积,通过设置不同的膨胀率将多尺度下的信息融合,解决遥感图像背景复杂的问题;使用EIOU边界框损失函数。在NWPU VHR-10数据集上验证,mAP达到94.5%,比原始YOLOv5s提高了1.2%。港口、车辆等小目标相应有1.3%和4.5%的提升。验证了算法对小目标识别、复杂背景识别的有效性。To solve the problems of small target,multi-scale and complex target background in remote sensing images,a Bottleneck Transformer target detection network was proposed.On the YOLOv5s model,the CNN+Transformer architecture was used to replace the C3 convolution operation in the last residual block,and the dilated convolution was used.The multi-scale information was fused by setting different expansion rates to solve the problem of complex remote sensing image background.The EIOU bounding box loss function was used.Verified on the NWPU VHR-10 dataset,the mAP reaches 94.5%,which is 1.2%higher than that of the original YOLOv5s.Small targets detection such as ports and vehicles have corresponding increases of 1.3%and 4.5%.The effectiveness of the algorithm for small target recognition and complex background recognition is verified.

关 键 词：遥感图像 目标检测 瓶颈变压器 特征融合 卷积神经网络 空洞卷积 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]