基于改进RT-DETR的遥感图像检测算法  

Remote Sensing Image Detection Algorithm Based on Improved RT-DETR

作　　者：白金燕 江涛[1] 魏玉梅 马珍 张琪 BAI Jinyan;JIANG Tao;WEI Yumei;MA Zhen;ZHANG Qi(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650500,China)

机构地区：[1]云南民族大学数学与计算机科学学院,云南昆明650500

出　　处：《无线电工程》2025年第2期334-342,共9页Radio Engineering

基　　金：国家自然科学基金(61363022)。

摘　　要：针对遥感图像目标排列紧密、背景复杂和小目标众多导致检测精度低的问题,提出了一种基于改进实时检测Transformer(Real-Time Detection Transformer, RT-DETR)的遥感图像检测算法。将Mosaic9数据增强应用到遥感数据中,丰富训练数据中场景和目标的组合,增强模型对不同环境下目标的识别能力。在主干网络中添加卷积块注意力模块(Convolutional Block Attention Module, CBAM),增强复杂背景下目标的关注度和图像特征提取能力,在模型中额外添加一个针对小目标的检测层,使小目标的细节特征更加突出,提升模型对小目标的检测能力。在DSTD舰船遥感数据集和NWPU VHR-10多类别遥感数据集上的实验结果显示,改进后的算法在交并比(Intersection over Union, IoU)阈值为0.5时,平均精度均值(mean Average Precision, mAP)分别达到了94.9%和94.5%,较原始RT-DETR算法分别提升了1%和1.3%,体现了改进算法在遥感图像检测上的有效性和通用性。To address issues of low detection accuracy in remote sensing images due to tightly arranged objects,complex backgrounds,and a high number of small objects,an improved detection algorithm based on the Real-Time Detection Transformer(RT-DETR)is proposed.This algorithm incorporates Mosaic9 data augmentation into the remote sensing data,enriching the combination of scenes and objects in training data and enhancing the model s ability to recognize objects under different environmental conditions.A Convolutional Block Attention Module(CBAM)is added to the backbone network to increase focus on objects against complex backgrounds and improve image feature extraction capabilities.Additionally,a specialized detection layer for small objects is integrated into the model,highlighting the detail features of small objects and enhancing small object detection capabilities.Experimental results on the DSTD maritime remote sensing dataset and the NWPU VHR-10 multi-category remote sensing dataset demonstrate that the improved algorithm achieves mean Average Precision(mAP)values of 94.9%and 94.5%respectively at an Intersection over Union(IoU)threshold of 0.5,showing improvements of 1%and 1.3%respectively over the original RT-DETR algorithm.These results confirm the effectiveness and versatility of the improved algorithm in remote sensing image detection.

关 键 词：目标检测 遥感图像 实时检测Transformer 卷积块注意力模块 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]