Instance segmentation method and device for remote sensing image road small target

By combining the DMH-YOLO model with the residual learning mechanism of Darknet-53 and C2f modules, and employing dual-stream feature pyramid fusion and multi-domain attention feature fusion, the problem of difficult segmentation of small targets in remote sensing images is solved, achieving higher precision instance segmentation results.

CN119027667BActive Publication Date: 2026-06-30SHANGHAI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SHANGHAI UNIV
Filing Date
2024-09-02
Publication Date
2026-06-30

AI Technical Summary

Technical Problem

In remote sensing images, small targets such as intersections and pedestrian crossings have limited visualization information, making it difficult to extract features and susceptible to interference from the external environment. This makes it difficult for instance segmentation models to accurately identify and segment them.

Method used

The DMH-YOLO model is used for instance segmentation. The residual learning mechanism of the Darknet-53 framework and C2f module is combined. Multi-resolution features are extracted through mix-pooling and soft-pooling. The backbone network features are fused using a dual-stream feature pyramid. The segmentation accuracy is improved by a multi-domain attention feature fusion module with spatial and frequency domain branches.

Benefits of technology

It improves the ability to capture information about small targets, enhances the network's segmentation accuracy for small road targets, and achieves more accurate instance segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027667B_ABST
    Figure CN119027667B_ABST
Patent Text Reader

Abstract

This application relates to a method and apparatus for segmenting road target instances in remote sensing images. The method includes: acquiring remote sensing images of the target to be detected by a remote sensing satellite; preprocessing the remote sensing images through an input module; extracting features from the preprocessed image using a backbone network; fusing the extracted multi-resolution features using a dual-stream feature pyramid in the neck region; dividing the head region into two branches: a detection head and a prototype; and performing category prediction and instance segmentation refinement tasks on the image entering the head region; outputting the high-resolution remote sensing image road target instance segmentation result; using a DMH-YOLO model to perform instance segmentation on the target remote sensing image to determine the target objects in the target remote sensing image; and using the Darknet-53 architecture framework in the backbone network and incorporating a C2f module to implement a residual learning mechanism. This invention can accurately identify and segment small target objects.
Need to check novelty before this filing date? Find Prior Art