Scientific Image Forgery Detection — Kaggle Competition

Participated in the ongoing Kaggle competition on Copy-Move Forgery Detection in Scientific Images, aimed at identifying manipulated biomedical figures that can compromise research integrity.

For this challenge, I developed a ResNet50 + U-Net hybrid segmentation model using PyTorch, designed to detect and segment forged regions at the pixel level. My approach combines Dice and Focal losses for balanced training, WeightedRandomSampling to oversample forged images, and Test-Time Augmentation (TTA) to improve prediction robustness.

Achieved an initial score of 0.303 on the public leaderboard. I’m continuing to experiment with architecture tuning, learning rate schedules, and other loss functions to further enhance performance and generalization.