Related Experiment Video
Updated: Aug 14, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
MDHANet: Rethinking HOG as a Prior in Dense Networks with Self-Attention for Hyperspectral Image Classification
Hongwei Zhang1, Yuanyuan Gui2, Junjie Mou1
1School of Artificial Intelligence, China University of Mining and Technology-Beijing, Beijing 100083, China.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
This study introduces a new Multiscale Dense HOG-driven Self-Attention Network (MDHANet) for hyperspectral image classification. MDHANet effectively addresses gradient vanishing and information loss, significantly improving classification accuracy by integrating Histogram of Oriented Gradients (HOG) prior knowledge.
Area of Science:
- Remote Sensing
- Computer Vision
- Machine Learning
Background:
- Convolutional neural networks (CNNs) excel in hyperspectral image (HSI) classification but face challenges with increasing depth, including gradient vanishing and information loss.
- Existing HSI classification methods often focus on novel architectures, neglecting the explicit gradient prior information within the data, which is crucial for complex scenarios.
Purpose of the Study:
- To propose a novel Multiscale Dense HOG-driven Self-Attention Network (MDHANet) for enhanced HSI classification.
- To leverage Histogram of Oriented Gradients (HOG) as prior knowledge to overcome limitations of traditional CNNs in HSI classification.
Main Methods:
- MDHANet utilizes three densely connected branches at different scales for multi-level feature representation and reuse.
- A dynamic HOG-driven self-attention module integrates HOG prior knowledge for feature rearrangement and a dual-branch self-attention mechanism to learn spatial information effectively.
Main Results:
- The proposed MDHANet method demonstrated superior performance over state-of-the-art approaches on four HSI datasets (IP, LK, HH, HC).
- Achieved high overall accuracies (OA) of 99.13%, 97.32%, 99.02%, and 99.24% respectively, with significant improvements over competing methods.
Conclusions:
- MDHANet effectively mitigates gradient vanishing and information loss issues in deep CNNs for HSI classification.
- The integration of HOG prior knowledge and a novel self-attention mechanism enhances the extraction of discriminative spatial-structure features, leading to improved classification performance.
