Remote Sensing Image Segmentation Method based on HRNET
Cheng Zhi, Daocai Fu
Abstract
High-Resoultion Net(HRNet) maintains high-resolution feature maps during the deep feature extraction process, and performs well in many areas of computer vision. In this paper, HRNet is applied to the road extraction and building extraction tasks of remote sensing images, and the two loss functions of DICE loss and BCE (binary crossentroy) loss can be used to solve the problem of samples imbalance. When there are only two classes of segmentation objects, HRNet performs better than several methods used in remote sensing image segmentation area. Experiments on road and building datasets shows that HRNet has strong practicality in remote sensing image segmentation.
Topics & Concepts
Computer scienceArtificial intelligenceSegmentationImage segmentationDiceComputer visionFeature extractionProcess (computing)Feature (linguistics)Scale-space segmentationImage (mathematics)Segmentation-based object categorizationRemote sensingImage resolutionPattern recognition (psychology)GeographyMathematicsPhilosophyLinguisticsGeometryOperating systemAutomated Road and Building ExtractionRemote Sensing and LiDAR ApplicationsAdvanced Neural Network Applications