DeepLense GSoC Assignment

This repository contains my implementation of the DeepLense GSoC 2026 evaluation tasks, including both classification and image super-resolution problems on gravitational lensing datasets.

---

Trained models can be found at: Google Drive - Model Checkpoints

---

Project Structure

  comman_test/
  │
  ├── other_experments/
  │ ├── image_classificationefficientnet_b3(300).ipynb
  │ ├── image_classificationefficientnet_b3(300)scheduld_learning_rate.ipynb
  │
  ├── comman_test.ipynb
  │
  project_task/
  │
  ├── task_1/
  │ ├── Task_VI_A.ipynb
  │
  ├── task_2/
  │ ├── outputs/
  │ │ ├── sample_visualizations.png
  │ │ ├── Tranning_Historys.png
  │ │
  │ ├── Task_VI_B.ipynb
  │
  README.md

---

Common Test I: Multi-Class Classification

Dataset and Classes

The dataset consists of three types of gravitational lensing images:

• No Substructure
• Sphere Substructure
• Vortex Substructure

---

Preprocessing Approach

• Resized images from 150×150 → 224×224
• Applied data augmentation:
  • Random Horizontal Flip (p = 0.5)
  • Random Vertical Flip (p = 0.5)
  • Random Rotation (0°–360°)
• Improved generalization using augmentation techniques

---

Model Architecture

Two state-of-the-art models were explored:

1. ConvNeXt-Base (Best Model)
2. EfficientNet-B3

Training Configuration (ConvNeXt-Base)

• Pretrained weights used
• Dropout: 0.3
• Loss: CrossEntropyLoss (label smoothing = 0.1)
• Optimizer: AdamW (lr = 1e-4, weight_decay = 1e-2)
• Scheduler: CosineAnnealingLR
• Epochs: 20

---

Results

Model	No Substructure	Sphere	Vortex
ConvNeXt-Base	1.00	0.99	1.00
EfficientNet-B3 (300×300)	0.98	0.95	0.98
EfficientNet-B3 + Scheduler	0.98	0.96	0.98

---

Classification Results

Training History (Classification)

---

Task VI.A: Super-Resolution on Simulated Data

Dataset

• Simulated strong lensing images
• Paired Low Resolution (LR) and High Resolution (HR)
• No substructure

---

Preprocessing

• LR: 75×75 → 64×64 (center crop)
• HR: 150×150 → 128×128
• Ensures compatibility with Swin Transformer
• Data split: 90% train / 10% validation

---

Model: Swin2SR (Transformer-Based SR)

• Transformer-based super-resolution model
• PixelShuffle upsampling for sharp reconstruction

Training Configuration

• Loss: L1 Loss
• Optimizer: AdamW (lr = 2e-4)
• Scheduler: CosineAnnealingLR
• Epochs: 20

---

Validation Results (1000 samples)

Metric	Value
MSE	0.00007577
SSIM	0.9751
PSNR	41.2477 dB

---

Training History (Task VI.A)

Sample Output (Task VI.A)

---

Task VI.B: Super-Resolution on Real Telescope Data

Dataset

• 300 real image pairs
• Sources: HSC & HST telescopes
• Limited dataset size

---

Preprocessing & Augmentation

• Min-Max normalization
• Data augmentation:
  • Rotation (90°, 180°, 270°)
  • Horizontal & Vertical Flip
  • Noise Injection

---

Model: EDSR (Best for Real Data)

• Residual-based SR architecture
• Effective for small datasets

---

Results

Bicubic Baseline

Metric	Value
MSE	0.005888
PSNR	23.58 dB
SSIM	0.3724

---

Final Model (EDSR)

Metric	Value
MSE	0.000683
PSNR	35.6530 dB
SSIM	0.8853

---

Training History (Task VI.B)

Super-Resolution Visualization (Task VI.B)

---

Author

Imrankhan