An end-to-end remote sensing image semantic segmentation method based on double tile input

By employing an end-to-end remote sensing image semantic segmentation method based on dual-pattern input, and utilizing patch pairs in the training set and majority voting, the problem of insufficient feature representation and computational efficiency in remote sensing image semantic segmentation is solved. This method achieves efficient and stable diverse classification results, improving classification accuracy and scalability.

CN122135032APending Publication Date: 2026-06-02LANZHOU JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
LANZHOU JIAOTONG UNIV
Filing Date
2026-03-12
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing remote sensing image semantic segmentation methods suffer from insufficient feature representation capabilities and computational efficiency in large-scale, multi-temporal, and high-resolution images, making it difficult to generate diverse classification results. Furthermore, multi-model ensemble methods have low computational efficiency and high storage management costs.

Method used

An end-to-end semantic segmentation method for remote sensing images based on dual-pattern input is adopted. The remote sensing image is divided into multiple patches. A convolutional neural network model is trained on the training set using the patches. The final category label of each pixel is determined by majority voting, thereby achieving the integration of diverse classification results.

Benefits of technology

Without requiring training multiple models, it improves classification accuracy and efficiency, ensures the model's speed, stability, and versatility, enhances the discriminative power between different categories, and expands the classification space.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122135032A_ABST
    Figure CN122135032A_ABST
Patent Text Reader

Abstract

This application discloses an end-to-end remote sensing image semantic segmentation method based on dual-tile input, comprising: dividing the remote sensing image into n tiles to construct a sample set; pairing each tile in the sample set to obtain n×n tile pairs, constructing a tile pair training set; training a convolutional neural network model using the tile pair training set to obtain an image semantic segmentation model; dividing the remote sensing image to be classified into multiple tiles, randomly selecting two tiles for pairing to construct a tile pair set, and determining the predicted combined label for each tile pair using the image semantic segmentation model; decomposing the predicted combined label to obtain the predicted category label of the original tile, thereby determining the diversity classification result of each pixel in the remote sensing image to be classified; determining the final category label of each pixel using a majority voting method, traversing all pixels to obtain the final semantic segmentation result of the remote sensing image to be classified. This application can ensure the speed and stability of end-to-end ensemble classification.
Need to check novelty before this filing date? Find Prior Art