This project involves two resources: Auto Labeled Image Splicing (ALIS) dataset and Forgery Segmentation Network (F-SegNet).
ALIS dataset is an image splicing verification dataset constructed based on ImageNet, COCO, and Crowd Human datasets. It comprises 224,388 spliced images with their corresponding ground truth masks.
F-SegNet can be used to detect image splicing forgeries. It smartly converts an image verification problem into forgery instance segmentation problem. By combining Mask R-CNN, Error Level Analysis, Normal Map Analysis, and MoCo v2 (a self-supervised learning method), F-SegNet can effectively locate spliced instances in a forged image.
ALIS dataset download: ALIS dataset
F-SegNet pretrained weights (after 500,000 iterations of training): F-SegNet final weights.
Install requirements:
pip install requirements.txt
To train F-SegNet, run the following code:
python train.py -d <string> -an <string> -i <int> -c <bool> -w <string> -o <string>
where:
-d is path to image directory
-an is path to COCO annotation file
-i is training iteration count
-c is continue training, which is set to True by default
-w is path to pretrained weights
-o is output directory path
Note:
If "continue training" is set to True, -w will be discarded and F-SegNet will continue training based on the last checkpoin in -o output directory.
To check training log, use tensorboard with the following code:
tensorboard --logdir <str>
where --logdir is the output directory path. It should be same with -o mentioned in Step 2.
To evalualte F-SegNet, use the following code:
python evaluate.py -d <string> -an <string> -w <string> -t <float>
where:
-d is path to image directory
-an is path to COCO annotation file
-w is output directory path, which should be the same as -o mentioned in Step 2. F-SegNet will automatically read the ./model_final.pth in this directory.
-t is the score threshold, which is set to 0.5 by default
This project is developed based on Detectron 2 framework.