FG-MSCNet

Frequency-Guided Multi-Scale Composite Network for document image tampering localization.

This repository is prepared for paper submission and focuses on reproducibility.

Highlights

• Dual-stream input stem: RGB + SRM residual features.
• Wavelet-based encoder-decoder with DWT downsampling and IDWT upsampling.
• Frequency-guided spatial reduction attention (FG-SRA).
• Dual-domain gate bridge (spatial MHSA + learnable spectral filter).
• Multi-scale deep supervision with composite loss (BCE + Dice + Edge).

Repository Status

• Paper link: coming soon
• Pretrained weights: Google Drive
• Inference demo script: coming soon

Environment Setup

1. Clone the repo:

git clone https://github.com/<your-username>/FG-MSCNet.git
cd FG-MSCNet

2. Activate conda environment:

conda activate <your_env>

3. Install PyTorch first (matching your CUDA):

• https://pytorch.org/get-started/locally/

4. Install remaining dependencies:

pip install -r requirements.txt

Dataset Layout

Use the DocTamper dataset with this directory layout:

<data_dir>/
  DocTamperV1-TrainingSet/
  DocTamperV1-TestingSet/
  DocTamperV1-FCD/
  DocTamperV1-SCD/

LMDB keys expected by dataset.py:

• num-samples
• image-%09d
• label-%09d

Training

Important: train.py uses distributed APIs internally. Run it with torchrun even on 1 GPU.

Single GPU:

torchrun --standalone --nproc_per_node=1 train.py \
  --data_dir <path/to/data_dir> \
  --exp_name fg_mscnet_exp1 \
  --runs_dir ./runs \
  --img_size 512 \
  --batch_size 2 \
  --epochs 200 \
  --lr 1e-4

Multi-GPU (example: 4 GPUs):

torchrun --standalone --nnodes=1 --nproc_per_node=4 train.py \
  --data_dir <path/to/data_dir> \
  --exp_name fg_mscnet_4gpu \
  --runs_dir ./runs

Resume full training state:

torchrun --standalone --nproc_per_node=1 train.py \
  --data_dir <path/to/data_dir> \
  --resume runs/<timestamp>_<exp_name>/checkpoints/last_model.pth

Load weights for finetuning:

torchrun --standalone --nproc_per_node=1 train.py \
  --data_dir <path/to/data_dir> \
  --finetune <path/to/checkpoint.pth>

Outputs are saved to:

runs/<timestamp>_<exp_name>/
  logs/train.log
  tensorboard/
  checkpoints/best_model.pth
  checkpoints/last_model.pth

Evaluation

Important: eval.py also uses distributed collectives. Use torchrun.

torchrun --standalone --nproc_per_node=1 eval.py \
  --data_dir <path/to/data_dir> \
  --checkpoint runs/<timestamp>_<exp_name>/checkpoints/best_model.pth \
  --img_size 512 \
  --save_dir ./eval_results \
  --save_n 50

Optional flag:

--pred_as_heatmap

Evaluation behavior:

• Tests on fcd, scd, and test.
• Searches threshold from 0.1 to 0.9 (step 0.1) and reports the best F1 threshold.
• Prints a markdown-style metrics table:
  • F1, Precision, Recall, mIoU, Accuracy, AUC-ROC, AP.
• Saves visualization panels under eval_results/<dataset_name>/.

Model Summary

python look.py

Method Overview

Input RGB
  -> InputStem (RGB + SRM fusion)
  -> Encoder x4 (DWT + FG-SRA)
  -> DDG Bridge
  -> Decoder x4 (IDWT + skip + mask guidance)
  -> Final head
Outputs: [H, H/2, H/4, H/8]

Project Structure

FG-MSCNet/
  layer/
    fg_mscnet.py
    modules.py
    dct.py
    srm.py
    wavelet.py
  dataset.py
  loss.py
  train.py
  eval.py
  look.py
  requirements.txt
  LICENSE
  README.md

Results

Dataset	F1	Precision	Recall	mIoU	Pixel Acc	AUC-ROC	AP
FCD	0.8483	0.9009	0.8016	0.7366	0.9898	0.9867	0.8803
SCD	0.9184	0.8703	0.9721	0.8491	0.9985	0.9995	0.9332
TEST	0.9714	0.9697	0.9730	0.9443	0.9993	0.9998	0.9926

License

This project is released under the MIT License. See LICENSE.