基于改进Mask-RCNN的遥感影像建筑物提取

由于遥感影像中建筑物种类繁多且与周围环境信息混淆,传统方法难以实现建筑物的准确高效提取.本文提出了一种基于改进Mask-RCNN的建筑物自动提取方法,利用PyTorch深度学习框架搭建改进Mask-RCNN网络模型架构,在网络的设计中添加了路径聚合网络和特征增强功能,通过监督和迁移学习的方式对Inria航空影像标签数据集进行多线程迭代训练与模型优化学习,实现了建筑物的自动精确分割和提取.基于不同开源数据集,分别与SVM、FCN、U-net和Mask-RCNN等建筑物提取算法进行对比,实验表明,本文方法可以高效准确、高效地提取建筑物,对于同一个数据集,提取结果的mAP、mRecall、mPrecision和F1分数这4个评价指标均优于对比算法.

Building Extraction from Remote Sensing Image Based on Improved Mask-RCNN

In view of the variety of buildings, and the confusion with the surrounding environment in the remote sensing image, the traditional methods are difficult to extract the buildings efficiently and accurately. This study proposes a method of building automatic extraction method based on the improved Mask-RCNN. In the proposed method, an improved Mask-RCNN network model framework is constructed by using PyTorch deep learning framework, path aggregation network and feature enhancement function are added in the network design, and multi-threaded iterative training and model optimization learning are conducted for the Inria aerial image tag data set by means of supervision and migration learning, thus automatic and accurate segmentation and extraction of buildings are achieved. The proposed method is compared with SVM, FCN, U-net, Mask-RCNN, and other building extraction algorithms with different open-source datasets. The experimental results show that the proposed method has competitive performance, which can extract buildings more efficiently, accurately, and quickly in different open-source datasets, and the four evaluation indexes of mAP, mRecall, mPrecision, and F1 scores extracted in the same dataset are better than the compared algorithms.