Dual-Image Super-Resolution using EDSR (with Concatenation Fusion)

A deep learning-based Dual-Image Super-Resolution framework built using a modified EDSR (Enhanced Deep Super-Resolution) architecture. The model takes two temporally shifted low-resolution satellite images and fuses them to generate a high-resolution output. Fusion is achieved via simple channel-wise concatenation followed by deep residual refinement.

---

📖 Overview

This project targets the problem of enhancing the spatial resolution of satellite images captured at different times. Since high-res satellites are costly and limited, we leverage multiple low-resolution inputs captured over time to reconstruct the corresponding high-resolution image.

This approach is especially useful for missions like Proba-V, where the trade-off between spatial and temporal resolution limits imagery quality.

---

📝 Articles & Documentation

I’ve written a series of Medium blogs that explain this project step-by-step — from dataset preparation to building the model, training it, and visualizing results.

You can read the full journey here:

• 📄 Part 1: Dataset Creation & Preprocessing Covers how the Proba-V satellite dataset was normalized, filtered, and restructured for dual-image SR training.

• 🧠 Part 2: Building the Model Architecture (EDSR + Fusion) Explains the PyTorch model architecture including the use of EDSR blocks, concatenation fusion, and PixelShuffle.

• 📊 Part 3: Training, Evaluation (PSNR), and Inference Discusses training strategy, metrics, loss functions, and how to visualize and interpret the results.

  📖 All concepts are explained clearly with examples, visuals, and reasoning, so don’t worry if you’re new to this!

These articles are written in a beginner-friendly tone with technical clarity, so even if you’re just getting started with deep learning or super-resolution, you’ll find them easy to follow.

If you find the articles helpful, feel free to give them a follow or clap on Medium — it helps spread the word and supports open-source learning!

---

Let me know if you'd like to add image previews or article badges!

🗂️ Dataset

• Main dataset PROBAV
• Preprocessed and final dualsr dataset

This project uses a preprocessed version of the Proba-V Super-Resolution Dataset. The structure is as follows:

dual\_sr\_dataset/
├── train/
│   ├── low\_res/
│   │   ├── imgsetXXXX\_LR0.png
│   │   ├── imgsetXXXX\_LR1.png
│   └── high\_res/
│       ├── imgsetXXXX\_HR.png
├── test/
│   ├── low\_res/
│   └── high\_res/

Each sample consists of:

• LR0: Low-res image at time t
• LR1: Low-res image at time t+1
• HR: Ground truth high-res image

---

🧠 Model Architecture

The model combines dual inputs using channel-wise concatenation after feature extraction and processes the merged tensor through an EDSR-style backbone for refinement.

LR1        LR2
│          │
\[Conv]    \[Conv]
│          │
└─> \[Concat + Fusion Block] ──> \[EDSR Backbone] ──> \[Upsample] ──> SR Output

Components:

• InitialConvBlock: 3×3 conv + ReLU per LR input
• FusionBlock: 1×1 conv to reduce channel size
• EDSRBackbone: Stack of Residual Blocks (default: 8)
• UpsampleBlock: Conv → PixelShuffle to upscale

---

⚙️ Training Details

• Framework: PyTorch
• Input Size: LR = 64×64, HR = 256×256
• Scaling Factor: ×4
• Loss Function: L1 / Combined Loss (L1 + PSNR)
• Optimizer: Adam
• Learning Rate: 1e-4
• Epochs: 50
• Batch Size: 4

Model checkpoints are saved in the checkpoints/ directory.

---

📊 Evaluation Metrics

The performance is evaluated using:

• PSNR (Peak Signal-to-Noise Ratio)

Higher values indicate better reconstruction.

---

✅ Results

Example: Dual inputs, model output, and ground truth comparison.

🔍 Sample Output

Here’s an example of how the model enhances resolution:

Sample outputs and intermediate visualizations are stored in the sr_visualization/ directory.

---

📚 References

• EDSR Paper (CVPR 2017)
• Proba-V Dataset
• PixelShuffle Docs

---