Remote sensing image segmentation method and system based on lightweight UMFormer

By constructing the LEUMFormer model and combining it with DecoupleNet, MSAF, SCIA, and BEH modules, the computational redundancy and boundary loss problems of UMFormer in farmland remote sensing image segmentation are solved, achieving high-precision and lightweight farmland remote sensing image segmentation, which is suitable for resource-constrained edge platforms.

CN122434963APending Publication Date: 2026-07-21ANHUI UNIV
1 Cites 1 Cited by

Patent Information

Application Number
CN202610535160.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-22
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

The existing UMFormer model has problems in farmland remote sensing image segmentation, such as redundant computation of multi-scale attention modules, insufficient modeling of orientation-sensitive features, lack of adaptive balance between spatial and channel attention, and loss of boundary details due to deep downsampling, which makes it difficult to meet the requirements of real-time deployment on the edge.

Method used

The LEUMFormer model is constructed, using DecoupleNet as a lightweight encoding network. A multi-scale attention fusion module (MSAF) and a global-local decoding network are introduced, including a decoder GL-VSSBlock, a spatial and channel adaptive integration module (SCIA), and a boundary enhancement module (BEH). Through multi-scale feature fusion, orientation awareness, and boundary enhancement, the network structure is optimized to improve segmentation accuracy and efficiency.

Benefits of technology

It significantly improves the accuracy and boundary continuity of farmland remote sensing image segmentation, reduces the computational load and parameter count of the model, meets the requirements for real-time deployment on the edge, and provides more complete and continuous segmentation results that are adaptable to different lighting and terrain conditions, thus possessing strong engineering practical value.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to the technical field of remote sensing image processing, and particularly discloses a remote sensing image segmentation method, system and product based on lightweight UMFormer. According to the method, on the basis of the UMFormer with light weight and moderate precision, a two-stage optimization strategy of firstly improving the segmentation precision and then balancing the light weight efficiency is adopted, and high-precision and light-weight segmentation of the remote sensing image is realized. According to the constructed LEUMFormer model, a DecoupleNet lightweight backbone network is adopted, channel redundancy is inhibited, and model lightweight is realized on the basis of precision improvement. Through an MSAF module, enhanced extraction of farmland multi-scale features and direction sensitive structures is realized with linear complexity. In the decoding stage, an SCIA module is introduced, and space structure and spectrum channel characteristics are dynamically fused through a double-branch parallel structure and self-adaptive gating. In addition, a BEH module is constructed, fine boundaries are explicitly strengthened with extremely low calculation cost, and the problems of fuzzy segmentation and fracture are solved.
Need to check novelty before this filing date? Find Prior Art