Related Experiment Video
Updated: Jun 8, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Ultralightweight progressive feature disentanglement and recomposition network for hyperspectral image
Delong Kong1, Shichao Zhang1, Xiang Yu1
1Remote Sensing Information and Digital Earth Center, College of Computer Science and Technology, Qingdao University, Qingdao, 266071, China.
Summary
This study introduces the Ultralightweight Progressive Feature Disentanglement and Recomposition Network (ULite-FDRNet) for efficient hyperspectral image classification. ULite-FDRNet achieves high accuracy with significantly fewer parameters and faster inference, making it suitable for resource-constrained applications.
Area of Science:
- Computer Vision
- Machine Learning
- Remote Sensing
Background:
- Deep Learning-based Hyperspectral Image Classification (DL-HSIC) requires effective spatial-spectral representations.
- Existing DL-HSIC models often have high parameter counts and inference latency, hindering practical deployment.
- There is a need for efficient DL-HSIC models that balance performance and computational cost.
Purpose of the Study:
- To propose a novel Ultralightweight Progressive Feature Disentanglement and Recomposition Network (ULite-FDRNet) for hyperspectral image classification.
- To achieve a superior trade-off between representation capability, parameter efficiency, and classification performance.
- To enable practical deployment of DL-HSIC in resource-constrained environments.
Main Methods:
- The study introduces the Feature Disentanglement and Recomposition (FDR) paradigm.
- Key modules include FDRConv2D/3D for local feature extraction, ScaleFDR3D for multiscale feature diversity, and TriSFDR for synergistic attention.
- These components facilitate progressive feature learning from shallow to global levels.
Main Results:
- ULite-FDRNet demonstrated superior classification accuracy across four HSI benchmarks.
- The model achieved high Overall Accuracies (OA) ranging from 97.47% to 99.57%.
- ULite-FDRNet utilizes significantly fewer parameters (0.86K to 3.51K) compared to existing methods, enabling faster inference.
Conclusions:
- ULite-FDRNet offers an efficient and effective solution for hyperspectral image classification.
- The proposed network achieves high accuracy with remarkable parameter efficiency and speed.
- ULite-FDRNet is well-suited for practical deployment in scenarios with limited computational resources.
