MHNet: a hybrid network for high-resolution remote sensing image semantic segmentation based on multiscale feature fusion

Abstract

Semantic segmentation of high-resolution remote sensing images (HRSIs) presents significant challenges, such as discrete object distributions, diverse scales, and class imbalance, which lead to problems like blurred boundary segmentation and insufficient global semantic associations. Although traditional convolutional neural networks are excellent in local feature extraction, their inherent structure limits the modeling of long-range dependencies. Transformer can model global context, but the quadratic complexity of the self-attention mechanism leads to high computational costs when dealing with HRSIs. Therefore, this manuscript proposes a novel encoder-decoder network called Multiscale Hybrid Network (MHNet) to effectively improve the segmentation performance of HRSIs through multiscale feature fusion, global context modeling, and boundary detail optimization. Specifically, in the encoder, the Neighborhood Feature Fusion (NFF) module is designed to fuse neighboring layer features, and aggregate low-level details and high-level semantics by channel and spatial attention. For the decoder, the Multiscale Refinement Enhanced Transformer Block (MRETB) and the Multiscale Refinement Attention Fusion (MRAF) module are proposed. Among them, MRETB uses the Multiscale Refinement Enhancement (MSRE) module to extract multiscale features and enhance boundary information, and the Window-based Efficient Multi-Head Self-Attention Mechanism (W-EMSA) to model long-range dependencies. MRAF further integrates the multiscale global context output by MRETB through integrating multilayer features and optimizing boundary details. The performance of MHNet is verified by experiments conducted on three public remote sensing image datasets.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata

Altmetrics