A hyperspectral image classification method based on a multi-stream feature fusion network

By using a multi-stream feature fusion network, combined with multi-level wavelet transform and guided feature fusion module, the problem of insufficient feature extraction in hyperspectral image classification is solved, achieving higher classification accuracy and robustness, and improving the overall performance of the model.

CN120976647BActive Publication Date: 2026-06-09湖南省第二测绘院 +1
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
湖南省第二测绘院
Filing Date
2025-08-13
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

In hyperspectral image classification, existing deep learning methods struggle to effectively extract information at different levels and scales with limited training samples, resulting in insufficient classification accuracy and inadequate information transfer between feature maps.

Method used

A multi-stream feature fusion network is adopted, which includes N+1 branches and a multi-level wavelet transform module. Combined with a guided feature fusion module, features are extracted through convolutional pooling and wavelet transform at different scales. The frequency domain attention module is used to enhance feature representation. At different stages of feature extraction, the mean max bi-branch pooling module and the multi-level wavelet transform module are introduced to capture local and global contextual information.

Benefits of technology

It improves the accuracy and robustness of hyperspectral image classification, enhances the model's generalization ability under different conditions, enables it to better learn the intrinsic structure of data, and improves feature recognition accuracy and overall performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120976647B_ABST
    Figure CN120976647B_ABST
Patent Text Reader

Abstract

This application discloses a hyperspectral image classification method based on a multi-stream feature fusion network. The method includes: acquiring a hyperspectral image dataset, preprocessing it to obtain training samples and test samples; inputting the training samples into a multi-stream feature fusion network for training, wherein the multi-stream feature fusion network contains N+1 branches and N sub-streams, each branch consisting of a first convolutional pooling layer, an average maximum bi-branch module, a second convolutional pooling layer, and a multi-level wavelet transform module; the output features of the five branches are integrated into a guided feature fusion module and then processed through a fully connected layer to obtain feature representations; the trained multi-stream feature fusion network is tested using the test samples, and the tested multi-stream feature fusion network is used to classify hyperspectral images. This method can fuse more features, avoids information loss during the classification process, and achieves higher classification accuracy.
Need to check novelty before this filing date? Find Prior Art