改进Faster RCNN的遥感图像建筑物检测  

Improved Faster RCNN for Remote Sensing Image Building Detection

作　　者：丁飞 石颉 袁晨翔 DING Fei;SHI Jie;YUAN Chenxiang(School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009)

机构地区：[1]苏州科技大学电子与信息工程学院,苏州215009

出　　处：《计算机与数字工程》2024年第9期2798-2803,2841,共7页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:62073231)资助。

摘　　要：为解决现有检测算法对遥感建筑图像检测率低,定位不准确等问题,提出一种改进Faster RCNN的遥感图像建筑物检测方法。在特征提取网络ResNet50上加入注意力机制SE模块,提升网络效率;将感兴趣区域校准(ROI Align)替换原有的感兴趣区域池化(ROI Pooling),消除两次量化带来的误差;利用GA-K-means算法获得更适合数据集的锚框。在WHU building dataset数据集上进行仿真实验,结果显示,改进后的Faster RCNN算法对遥感图像建筑检测的平均精度为92.21%,相较于原算法提升了10.79%,验证了改进方法的有效性。To solve the problems of low detection rate and inaccurate localization of remote sensing building images by existing detection algorithms,a remote sensing image building detection method with improved Faster RCNN is proposed.The attention mechanism which called improved sequeeze and excitation is added to the feature extraction network ResNet50 to improve the network efficiency.The original ROI Pooling is replaced by the ROI Align to eliminate the error caused by two quantizations.The GA-K-means algorithm is used to obtain more suitable anchor boxes.Simulation experiments on WHU building dataset and the results show that the average accuracy of the improved Faster RCNN algorithm for remote sensing image building detection is 92.21%,which is 10.79% higher than the original algorithm,and the effectiveness of the improved method is verified.

关 键 词：Faster RCNN 注意力机制 ROI Align GA-K-means 建筑物检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]