GoldNet

GoldNet is a public benchmark dataset and code release for visual authentication of genuine versus counterfeit gold items from ordinary smartphone photographs. It accompanies the paper:

GoldFormer: A Texture-Aware Vision Transformer-Based Algorithm for Detecting Near-Identical Images, Z. Raisi, Algorithms, 2026, 19(7), 530. DOI: 10.3390/a19070530. Open access (CC BY 4.0).

The task is fine-grained: high-quality counterfeits closely replicate the surface texture, hallmark engravings, color, and geometry of genuine pieces, so the two classes are near-identical to the eye. On a blind subset, trained gold-trading experts reached 89.80% accuracy, which sets the human baseline the models are compared against.

Dataset

• 2,127 images of physical gold items, one image per item (no item is photographed more than once).
• 1,044 authentic (real) and 1,083 counterfeit (fake), a near-balanced split (49.1% / 50.9%).
• Captured with several consumer smartphones under varied real-world conditions (daylight, indoor, and low-light; a range of angles, distances, and backgrounds), with no specialist imaging hardware.
• Items originate primarily from Iran and the wider Persian Gulf market.

Layout

gold/
  real/    # authentic items   (r_img_001.jpg ...)   1,044 images
  fake/    # counterfeit items (f_img_001.jpg ...)   1,083 images
  pairs/   # matched authentic/counterfeit examples used in the paper figures

Because each image is a distinct physical item, an image-level train/validation split is also an item-level split: no item can appear in more than one fold, so the cross-validation results carry no item-level leakage.

Code

code/        # training and evaluation pipeline (PyTorch)
weights/     # pretrained backbone checkpoints and trained model weights

The canonical evaluation uses 5-fold stratified cross-validation, AdamW, AMP (bfloat16), and a freeze-then-unfreeze fine-tuning schedule.

# environment (CUDA 11.8 build of PyTorch)
python -m venv .venv && source .venv/bin/activate
pip install torch==2.4.1+cu118 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# run the full benchmark (ResNet / VGG / ViT / Swin / GoldFormer ...)
python run_experiments.py

# evaluate only the 2025 backbones (iFormer-S, OverLoCK-XT)
python run_new_models.py

# classical SVM baselines (LBP, Haralick)
python run_svm_only.py

models.py includes the GoldFormer/TAAG architecture. To reproduce the published model from the released checkpoint:

import torch
from models import build_model

model = build_model("goldformer")
state = torch.load("weights/GoldFormer_best.pth", map_location="cpu", weights_only=True)
model.load_state_dict(state)   # strict — exact match with the released checkpoint
model.eval()
logits, gamma = model(images)  # gamma = TAAG gate activations, for interpretability

Results (5-fold cross-validation, matched 224×224 resolution)

Model	Accuracy (%)	F1
Human experts (baseline)	89.80	--
ResNet-101	92.29 ± 1.01	0.9228
Swin-T	93.65 ± 0.67	0.9365
ViT-B/16	94.31 ± 0.94	0.9431
Soft-voting ensemble	94.92	0.9492
GoldFormer (ours)	95.02 ± 0.75	0.9502

GoldFormer is the best single model and beats the ensemble; it is statistically tied with the strongest individual backbone, ViT-B/16 (paired McNemar p = 0.228), and significantly beats its own Swin-T backbone (p = 0.014) while using half ViT-B/16's FLOPs (8.6 vs 16.9 GFLOPs) and fewer parameters (54.3M vs 86.6M). Its contribution is competitive accuracy together with built-in, attribution-free texture-gate interpretability.

Citation

@article{raisi2026goldformer,
  title   = {GoldFormer: A Texture-Aware Vision Transformer-Based Algorithm
             for Detecting Near-Identical Images},
  author  = {Raisi, Zobeir},
  journal = {Algorithms},
  volume  = {19},
  number  = {7},
  pages   = {530},
  year    = {2026},
  doi     = {10.3390/a19070530}
}

This repository is archived on Zenodo: 10.5281/zenodo.21123708.

License

The code and scripts in this repository are released under the MIT License.

The dataset (gold/ directory) is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You are free to use, share, and adapt the data for any purpose, provided you give appropriate credit and cite the paper above.

Contact

Zobeir Raisi, Chabahar Maritime University, zobeir.raisi@cmu.ac.ir