基于多任务学习的复杂城市遥感图像道路提取  

Road Extraction from Complex Urban Remote Sensing Images Based on Multi-task Learning

作　　者：王坤阳 刘洋 业宁[1] 张凯 WANG Kunyang;LIU Yang;YE Ning;ZHANG Kai(School of Information Science and Technology,,Nanjing Forestry University,Nanjing 210037,China;School of Instrument Science and Engineering,Southeast University,Nanjing 210018,China)

机构地区：[1]南京林业大学信息科学技术学院,南京210037 [2]东南大学仪器科学与工程学院,南京210018

出　　处：《计算机科学》2024年第S02期408-415,共8页Computer Science

基　　金：未来网络科研基金项目(FNSRFP-2021-YB-17)。

摘　　要：提出一种新的遥感图像道路提取框架,旨在利用从道路边缘检测中获得的知识来提高道路提取的准确性。研究中引入了一个融合多尺度信息和视觉注意力机制的多尺度视觉注意力模块,并构建了一个级联特征融合模块以集成网络在不同尺度上的预测结果。在此基础上,构建了一个包含编码器和解码器的多尺度视觉注意网络(MSVANet)。同时,提出一个多任务学习框架,该框架结合了MSVANet,并采用粒子群优化算法对多任务学习框架的两个学习率超参数的自动选取进行优化。RNBD数据集的训练和测试结果表明,所提方法在各种分割精度指标和泛化能力方面均优于其他道路提取方法。In this paper,we propose a new framework for road extraction from remote sensing images that aims to utilize the knowledge gained from road edge detection to improve the accuracy of road extraction.A multi-scale visual attention module that fuses multi-scale information and visual attention mechanisms is introduced in the study,and a cascading feature fusion module is constructed to integrate the network’s prediction results at different scales.Based on this,we construct a multiscale visual attention network(MSVANet)containing encoders and decoders.A multi-task learning framework that incorporates the MSVANet is also proposed,and a particle swarm optimization algorithm(PSO)is used to optimize the automatic selection of the two learning rate hyperparameters of the multi-task learning framework.The training and testing results on the RNBD dataset show that the proposed method outperforms other road extraction methods in terms of various segmentation accuracy metrics and generalization ability.

关 键 词：深度学习 遥感图像道路提取 多任务学习 多尺度视觉注意力网络(MSVANet) 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]