---
license: apache-2.0
language:
- en
pipeline_tag: image-segmentation
tags:
- manipulation
- forgery
- image
- cnn
- transformer
- residual-noise
- efficientnet
- swin
- localization
---
# NGIML Model Card

## Inference

NGIML performs single-image forgery localization from a pretrained checkpoint and an input RGB image.

## Checkpoints

Pretrained checkpoints are hosted on Hugging Face:

[juhenes/ngiml](https://huggingface.co/juhenes/ngiml)

Available checkpoints:

- `casia-effnet.pt`
- `casia-effnet+noise.pt`
- `casia-effnet+swin.pt`
- `casia-full.pt`
- `casia-swin.pt`
- `casia-swin+noise.pt`

## Run Inference

### Recommended: Google Colab

The easiest way to test the model is through Colab:

- Google Colab: [Open `infer.ipynb` in Colab](https://colab.research.google.com/github/juhenes/ngiml-infer/blob/main/infer.ipynb)

This is the recommended path for quick testing because the notebook is already set up for checkpoint-based inference.

### Local CLI

If you want to run the project locally, use the repository files here:

- GitHub: [juhenes/ngiml-infer](https://github.com/juhenes/ngiml-infer)

Install the dependencies:

```bash
pip install -r requirements.txt
```

Run the CLI:

```bash
python predict.py --checkpoint /path/to/checkpoint.pt --image /path/to/image.png
```

Example:

```bash
python predict.py --checkpoint checkpoints_cache/casia-full.pt --image /path/to/image.png
```

If `--output-dir` is omitted, outputs are saved under `outputs/<image-stem>/`.

## Optional Arguments

- `--output-dir` to choose where outputs are saved
- `--threshold` to override the default binary threshold
- `--normalization-mode` to set `imagenet` or `zero_one`
- `--resize-max-side` to resize large images before preprocessing
- `--crop-size` to override the inference crop size
- `--device` to choose a device such as `cpu` or `cuda:0`

## Output Files

When an output directory is used, the runtime saves:

- `input_rgb.png`
- `preview_input_rgb.png`
- `preview_probability_map.png`
- `preview_binary_mask.png`
- `preview_overlay.png`
- `probability_map.png`
- `binary_mask.png`
- `overlay.png`
- `prediction.json`

`prediction.json` includes summary metadata such as the checkpoint path, threshold, normalization mode, device, and basic prediction statistics.

## References

1. Dong, J., Wang, W., and Tan, T. "CASIA Image Tampering Detection Evaluation Database." 2013 IEEE China Summit and International Conference on Signal and Information Processing, 2013. [DOI](https://doi.org/10.1109/chinasip.2013.6625374)

```bibtex
@inproceedings{Dong2013,
  doi = {10.1109/chinasip.2013.6625374},
  url = {https://doi.org/10.1109/chinasip.2013.6625374},
  year = {2013},
  month = jul,
  publisher = {{IEEE}},
  author = {Jing Dong and Wei Wang and Tieniu Tan},
  title = {{CASIA} Image Tampering Detection Evaluation Database},
  booktitle = {2013 {IEEE} China Summit and International Conference on Signal and Information Processing}
}
```

2. Pham, N. T., Lee, J.-W., Kwon, G.-R., and Park, C.-S. "Hybrid Image-Retrieval Method for Image-Splicing Validation." Symmetry, 11(1), 83, 2019.

```bibtex
@article{pham2019hybrid,
  title = {Hybrid Image-Retrieval Method for Image-Splicing Validation},
  author = {Pham, Nam Thanh and Lee, Jong-Weon and Kwon, Goo-Rak and Park, Chun-Su},
  journal = {Symmetry},
  volume = {11},
  number = {1},
  pages = {83},
  year = {2019},
  publisher = {Multidisciplinary Digital Publishing Institute}
}
```