ezzhamed/swin_v2

Scientific Image Forgery Detection (Swin V2 + Frequency Analysis)

Experiment Overview: swin_v2_frequency_optimization.ipynb

This experiment represents a significant upgrade from the baseline model, designed to tackle the subtleties of scientific image forgery. It introduces a Hybrid Architecture that fuses the spatial hierarchy of Swin Transformer V2 with forensic-specific enhancements.

Note: This experiment is based on the Recod.ai/LUC - Scientific Image Forgery Detection Kaggle competition.

Key Experimental Changes

1. Advanced Backbone Architecture

• Upgraded to Swin V2: Transitioned from Swin V1 to swin_v2_base_window12_384.
• Why? Swin V2 better handles high-resolution inputs and mitigates training instability (e.g., activation spikes) common in large-scale transformers.

2. Forensic Data Augmentation Pipeline

To simulate realistic forgery scenarios, we implemented a specialized augmentation strategy:

• JPEG Compression (Quality 60-100): Trains the model to detect double-compression artifacts.
• Gaussian Noise & Blur: Mimics sensor noise discrepancies between spliced regions.
• Random Erasing: Forces the model to learn robust features rather than relying on local textures.

3. Frequency Domain Analysis (DCT)

• Integrated Discrete Cosine Transform (DCT) visualization tools.
• Purpose: To reveal high-frequency anomalies and compression artifacts invisible to the human eye, acting as a secondary validation layer.

4. Robust Inference Strategy (TTA)

• Implemented Test Time Augmentation (TTA).
• Mechanism: Aggregates predictions from the original image + horizontal/vertical flips.
• Impact: Significantly reduces false negatives by ensuring consistent predictions across different orientations.

Results on Supplemental Dataset

This experiment yielded valid, perfect results on the challenging external validation set:

• Detection Rate: 100% (Correctly flagged 48 of 48 forged samples).
• F1-Score: 100% (Perfect classification at optimized threshold).
• Optimal Threshold: Identified 0.1 as the critical threshold for maximum sensitivity.