A remote sensing image instance segmentation method based on semi-supervised learning

By constructing a semi-supervised learning framework and adaptive enhancement of remote sensing images, and screening high-quality pseudo-labels, the problems of noise and imbalance in remote sensing images are solved, and low-cost, fine-grained remote sensing target perception and high-quality instance segmentation are achieved.

CN122223338APending Publication Date: 2026-06-16ANHUI UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ANHUI UNIV OF SCI & TECH
Filing Date
2026-05-12
Publication Date
2026-06-16

AI Technical Summary

Technical Problem

In existing semi-supervised instance segmentation methods, the mask pseudo-labels have a lot of noise near the target boundary. The noise affects the model's ability to perceive edge details. Furthermore, the imbalance between the foreground and background and the speckle noise in the remote sensing image increase the segmentation difficulty and labeling cost, making it difficult to achieve low-cost, fine-grained remote sensing target perception.

Method used

A semi-supervised learning framework is constructed, employing teacher and student networks. High-quality pseudo-labels are selected through preheating training, edge refinement perception module (ERP), and mask scoring branch (MPCF). Combined with remote sensing image adaptive enhancement (RSAE) to process foreground-background imbalance and speckle noise, low-cost fine-grained remote sensing target perception is achieved.

Benefits of technology

It reduces the cost of remote sensing image annotation, improves the model's ability to perceive targets in remote sensing images with fine granularity, enhances its adaptability to noise and imbalanced data, and achieves high-quality instance segmentation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122223338A_ABST
    Figure CN122223338A_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of image segmentation, and particularly relates to a remote sensing image instance segmentation method based on semi-supervised learning. The present application constructs a semi-supervised learning framework containing a teacher network and a student network, initializes the teacher network after the student network is made to have basic segmentation capability through preheating training; in the formal training stage, supervised and unsupervised training are executed in parallel, the pseudo-labels generated by the teacher network are screened in terms of class score and mask quality score by using a mask pseudo-label comprehensive screening mechanism in the unsupervised training, and the high-quality pseudo-labels screened are used to guide the student network to learn; meanwhile, remote sensing image adaptive enhancement operations are adopted, including imbalance challenge adaptive enhancement and SAR image challenge adaptive enhancement, so as to cope with the imbalance of foreground and background and the challenge of speckle noise. The present application has achieved instance segmentation performance superior to that of existing semi-supervised learning methods on both visible light and SAR remote sensing image datasets, and significantly reduces the dependence on pixel-level annotation.
Need to check novelty before this filing date? Find Prior Art