基于改进Mask RCNN算法的遥感建筑物检测  

Remote sensing building detection based on improved Mask RCNN algorithm

作　　者：熊风光[1,2] 董彪 张鑫[1] 刘欢乐 韩燮[1,2] 况立群[1,2] XIONG Feng-guang;DONG Biao;ZHANG Xin;LIU Huan-le;HAN Xie;KUANG Li-qun(College of Data Science and Technology,North University of China,Taiyuan 030051,China;Shanxi Province’s Vision Information Processing and Intelligent Robot Engineering Research Center,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学大数据学院,山西太原030051 [2]中北大学山西省视觉信息处理及智能机器人工程研究中心,山西太原030051

出　　处：《计算机工程与设计》2023年第1期218-223,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(62272426);山西省回国留学人员科研基金项目(2020-113);山西省科技成果转化引导专项基金项目(202104021301055);太原市专利导航基金项目(DH20210302017)。

摘　　要：为解决部分遥感建筑物因为自身形状的不规则,导致传统矩形识别框算法对该类检测目标分割效果差,难以精确定位的问题,提出一种改进的Mask RCNN检测算法。改进Mask RCNN的主干网络FPN网络,简化特征融合过程,有效避免语义信息丢失;改进Mask RCNN的RPN网络,针对识别框的重复计算,提升其运算效率,提高检出率;调节mask掩膜参数,提高分割效果。实验结果表明,改进Mask RCNN目标检测算法的检测精度和召回率达到了99.80%和97.88%,较原算法分别提高了1.54%和1.65%,有效优化了遥感领域不规则建筑物的检测问题。To solve the problems that the traditional rectangular recognition box algorithm has poor segmentation effect on this kind of detection target and it is difficult to accurately locate due to the irregular shape of some remote sensing buildings,an improved Mask RCNN detection algorithm was proposed.The backbone network FPN of Mask RCNN was improved to simplify the process of feature fusion and effectively avoid the loss of semantic information.The RPN network of Mask RCNN was improved to improve its operation efficiency and detection rate for the repeated calculation of the recognition box.The mask parameters were adjusted to improve the segmentation effect.Experimental results show that the detection accuracy and recall rate of the improved Mask RCNN target detection algorithm reach 99.80%and 97.88%,which are 1.54%and 1.65%higher than that of the original algorithm,respectively.The detection problem of irregular buildings in remote sensing field is effectively optimized.

关 键 词：目标检测 遥感 深度学习 图像分割 图像处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]