HSI-DMGASR (AAAI 2024)

Authors: Zhaoyang Wang, Dongyang Li, Mingyang Zhang†, Hao Luo, Maoguo Gong

This repository provides the official implementation of the AAAI 2024 paper:
"Enhancing Hyperspectral Images via Diffusion Model and Group-Autoencoder Super-Resolution Network"

The proposed model, HSI-DMGASR, integrates a Group-Autoencoder (GAE) with a Diffusion Model to tackle the challenges of hyperspectral image super-resolution (HSI-SR).

🔧 The code is based on prior works including SR3, SSPSR, GELIN, and MCNet, and is structured in two stages:

1. Training the Group-Autoencoder (GAE)
2. Joint training with the Diffusion Model for HSI Super-Resolution

---

🧱 Network Architecture

---

🧠 Motivation

Existing hyperspectral image (HSI) super-resolution (SR) methods struggle to effectively capture the complex spectral–spatial relationships and fine-grained visual details. In contrast, diffusion models have demonstrated powerful generative capabilities in modeling both high-level semantics and low-level structures. However, their direct application to HSI-SR faces challenges including convergence difficulty and high inference time.

To overcome these issues, we propose a Group-Autoencoder (GAE) that:

• Encodes high-dimensional HSI data into a compact latent space
• Facilitates stable diffusion model training
• Preserves spectral correlation
• Significantly accelerates inference

This synergy forms the DMGASR framework, achieving superior performance on both natural and remote sensing HSI datasets—both visually and quantitatively.

---

📊 Experimental Results

---

⚙️ Installation

Install required dependencies using:

pip install -r requirements.txt

---

🚀 Usage Instructions

Step 1️⃣: Train the Group-Autoencoder (GAE)

After setting the dataset paths, run:

python AE.py

---

Step 2️⃣: Train the Diffusion Model

Once the GAE is trained, load its weights and run:

python sr_gae.py

🔄 Note: You can toggle between training and inference modes in sr_gae.py.

---

⚙️ Configuration

The training configuration file is located at:

EHSI-DMGESR/config/sr_sr3_16_128.json

• Use this file to set:
  • Dataset paths
  • Model parameters
  • Training/inference settings

📁 Data Loading Methods

We support two ways of loading HSI datasets:

1. TrainsetFromFolder

• Follows the MCNet-style data pipeline
• Data is pre-processed using MATLAB and loaded from local folders

2. HSTrainingData / HSTestData

• Handles online preprocessing
• Offers greater flexibility and is suitable for end-to-end pipelines
• Check the function definitions for more usage details

---

📌 Notes

• The codebase is primarily built upon SR3, SSPSR, and MCNet frameworks.
• We thank the authors of these works for their valuable contributions to the community.

---