Skip to content
David Lonski
← Back to research

layer: archive · research log

Mask R-CNN — Image Forgery Detection

Kaggle competition entry · 2025

A deep learning pipeline for detecting forged scientific images, built for a Kaggle competition on real competition data.

  • Python
  • PyTorch
  • Torchvision
  • Mask R-CNN

what I did

  • 01Implemented Mask R-CNN in PyTorch and Torchvision for object detection and instance segmentation.
  • 02Worked with the Recod.ai/LUC Scientific Image Forgery Detection data: 2,751 forged images with masks and 2,377 authentic ones, from 256×256 up to 1912×2568 pixels.
  • 03Trained for about two dozen epochs, tracking box and mask loss on training and validation sets.
  • 04Validation loss leveled off early while training loss kept falling, a sign of overfitting and the next thing to fix.

from the work

Slide describing the dataset, with a forged microscopy image and its mask
The competition data: forged images, their masks and authentic images.
Ground-truth forgery boxes next to the model's predicted masks on two images
Ground truth (left) against the model's predicted masks (right).
Slide with training and validation loss curves for boxes and masks
Training and validation loss for boxes and masks across epochs.

next log

EduAlly →