🚀 Planetary Image Enhancement using Residual CNN (ResNet-Based)

This project presents a residual CNN-based image enhancement pipeline for planetary surfaces, particularly focused on Mars and crater regions. The model enhances satellite image quality using patch-based training, residual learning, and reflection padding to preserve texture and structural continuity while eliminating patch artifacts.

---

📂 Dataset Overview

The dataset is structured into raw planetary patches and their corresponding enhanced (ground truth) versions. Patches are of size 512×512.

🔹 Raw Input Patches (To be Enhanced)

• Patches_mars/
• Patches_crater/

🔹 Ground Truth (Enhanced) Patches

• Patches_enhanced_mars/
• Patches_enhanced_crater/
• Patches_enhanced_mars2/

These datasets are used to train the model to learn the transformation from raw to enhanced planetary imagery.

---

🗂️ Folder Structure

• Patches_mars/ # Raw Mars patches
• Patches_crater/ # Raw Crater patches
• Patches_enhanced_mars/ # Ground Truth enhanced Mars
• Patches_enhanced_crater/ # Ground Truth enhanced Crater
• Patches_enhanced_mars2/ # Additional enhanced Mars data
• Enhanced_mars_resnet/ # Output of ResNet-enhanced Mars
• Enhanced_crater_resnet/ # Output of ResNet-enhanced Crater
• Patches_mars13/ # Test image patches (mars13.jpg)
• Enhanced_mars13_resnet/ # Enhanced patches from test image
• mars13.jpg # Raw large image
• resnet_mars13.png # Final stitched enhanced output
• mars_enhancement_resnet.pth # Trained ResNet model
• Chart.png # Metrics visualization chart

---

🧠 Model Architecture – Residual CNN (ResNet)

This CNN model uses a deep residual learning approach to enhance planetary images patch-by-patch. Key components:

• Input: 512×512 RGB patches
• Reflection Padding: Prevents border artifacts during patch convolution
• Residual Blocks: 8 blocks with skip connections to retain features and ease optimization
• Batch Normalization & ReLU: Used after each convolution
• Final Output: 3-channel enhanced patch with sigmoid activation

The model is optimized using Mean Squared Error loss and the Adam optimizer with a learning rate of 1e-4. Trained for 50 epochs with batch size 6.

---

🔁 Enhancement Pipeline

1. Image Splitting

   • Large raw images (e.g., mars13.jpg) are split into non-overlapping 512×512 patches using split_image_smooth.

2. Patch Enhancement

   • Each patch is passed through the trained ResNet model. Output patches are normalized and saved.

3. Stitching with Gaussian Feathering

   • Enhanced patches are stitched into the full image using final_smooth_stitch with Gaussian weight masks to blend overlapping regions and avoid visible seams.

---

📊 No-Reference Image Quality Metrics

Image	BRISQUE ↓	NIQE ↓	PIQE ↓	Entropy ↑	SNR ↑	HVS Sharpness ↑
original	1.56	1.10	99.17	5.70	53.78	1.32
GT	0.60	0.56	17.55	6.79	39.55	173.82
unet	0.38	0.38	40.33	6.64	48.98	11.09
resnet	0.48	0.49	51.18	7.29	40.56	81.94
gan	0.61	0.62	12.77	6.78	29.66	282.74

🔺 ↑ Higher is better, 🔻 ↓ Lower is better
✅ ResNet achieves highest entropy, balanced SNR, and strong sharpness while maintaining low BRISQUE and NIQE values.
✅ GAN shows extremely high HVS sharpness but lower SNR.
✅ U-Net gives lowest BRISQUE but slightly lower entropy than ResNet.

---

📊 Chart Visualization

The comparison of metrics across models is illustrated below:

• Top Row (↑ Higher is Better):

  • Entropy: ResNet achieves the highest, indicating rich texture.
  • SNR: ResNet is competitive, balancing sharpness with noise control.
  • HVS Sharpness: GAN is highest; ResNet balances detail with naturalness.

• Bottom Row (↓ Lower is Better):

  • BRISQUE & NIQE: U-Net performs best, followed by ResNet.
  • PIQE: GAN achieves lowest distortion; ResNet performs moderately well.

These metrics confirm that ResNet maintains a strong trade-off between sharpness, texture richness, and natural appearance.

---

🧠 Interpretation & Highlights

• Entropy (7.29): Indicates that ResNet enhances surface texture more than all models.
• BRISQUE (0.48): Competitive natural image quality, close to U-Net.
• SNR (40.56): Shows ResNet balances sharpness and denoising well.
• HVS (81.94): Outperforms U-Net significantly in perceptual sharpness.
• Visual Quality: ResNet images show no tile lines, enhanced clarity, and improved contrast.

---

🧾 Conclusion

The ResNet-based CNN enhancement model effectively improves planetary imagery by:

• Enhancing surface textures and shadows on Martian landscapes
• Avoiding patch boundary artifacts using reflection padding and Gaussian blending
• Producing sharp, clear, and natural-looking full images
• Achieving high entropy and perceptual sharpness with low distortions

When compared to U-Net and GAN:

• ResNet provides the best texture enhancement (entropy)
• Balances naturalness, sharpness, and smooth transitions
• Excels in scenarios requiring both structure and realism

This makes it highly suitable for planetary science, cartography, and research visualization of extraterrestrial terrains.

---