--- license: cc-by-nc-4.0 library_name: pytorch pipeline_tag: image-segmentation tags: - document-forensics - tampering-localization - image-forensics --- # DiffNet DiffNet localises tampered regions in document images. Given a document image, it predicts a per-pixel tampering probability map and an image-level tampering probability. Paper: [arXiv:2606.22285](https://arxiv.org/abs/2606.22285) ยท Code: ## Weights One folder per training protocol of the paper. Each folder contains the weights and the training config. | Folder | Protocol | Inference input size | |---|---|---| | `doc_protocol/` | Doc Protocol: trained on the DocTamper training set (512x512); tested on the DocTamper test sets and on T-SROIE, OSTF, Tampered-IC13, RealTextManipulation | 512 | | `tdoc_protocol/` | Syn2Real-TDoc protocol: trained on TDoc-2.8M (1024x1024) from the `doc_protocol/` weights; tested on RTM, FindItAgain, FindIt | 1024 | ## Inference The model code is in the GitHub repository (installation details in its README): ```bash git clone https://github.com/Mohamed-Dhouib/DiffNet DiffNet cd DiffNet pip install -r requirements.txt # after torch and mmcv, see the README cd training python infer.py --weights /diffnet_doc_protocol.safetensors --input \ --output-dir --input-size 512 ``` Use `--input-size 1024` with the `tdoc_protocol/` weights. Inference also runs on CPU (`--device cpu`). Larger images are processed with non-overlapping windows of the input size, smaller ones are padded with white. For each image, `infer.py` writes `_prob.png` (probability x 255) and `_mask.png` (probability > `--threshold`, default 0.5, 0/255) and prints the image-level probability. Training and evaluation instructions: `README.md` and `REPRODUCIBILITY.md` of the GitHub repository. Results are reported in the paper. ## Citation ```bibtex @inproceedings{dhouib2026efficient, title={Efficient Document Tampering Localization with Multi-level Discrepancy Features and Unified DCT--Quantization Embedding}, author={Dhouib, Mohamed and Zhu, Ye and Vanier, Sonia and Shabou, Aymen}, booktitle={European Conference on Computer Vision}, pages={110--127}, year={2026}, organization={Springer} } ``` ## Licence The weights are released under CC BY-NC 4.0 (they are initialised from the ConvNeXt-V2 ImageNet checkpoints, which are CC BY-NC 4.0). The code on GitHub is MIT.