Lightweight Hybrid Siamese Neural Network

This project contains the Python implementation used to train, validate, and test the proposed lightweight hybrid Siamese neural network on a personal computer.

The trained models can be exported to TFLite format for deployment on resource-constrained edge devices.

Contents

• benchmark_energy.py - Measures GPU power consumption and energy per inference for all models.
• best_model_score.py - The scoring method follows the weighted normalization strategy proposed in Romano et al. [1], enabling fair comparison of models under energy-constrained edge-AI environments.
• blocks.py - Building blocks used in models.py (Conv, Depthwise Conv, Residual Block, Inverted Residual Block, SE block, etc.).
• data_loader.py - Dataset loader and label-pair generator for Siamese training.
• labels_vis.py - Visualizes positive and negative training pairs.
• main.py - Main training script.
• models.py - Network architectures (Ours, ResNet18, MobileNet family, and several custom experimental models).
• multiobjective_scores_result.csv - Multi-objective score of all models, computed using a combined Z-score and min-max normalization.
• my_metrics.py - Custom evaluation metrics (contrastive loss, Euclidean distance, ROC/EER computation, etc.).
• plot_all_model_db.py - Visualization of ROC/DET curves for all models across all datasets.
• plot_ours_all_db.py - Visualization of the proposed Ours model over all folds and datasets.
• plot_all_model_history.py - Plots training and validation accuracy/loss curves for every model.
• plot_utils.py - Utility functions for metric computation and plotting (confusion matrix, DET curve, EER point, etc.).
• requirements.txt - Python 3.9.2 dependency list.

Datasets

Three wrist-vein datasets were used for model training:

• NTUST-IB811 [2]: Collected using our imaging device https://ieee-dataport.org/documents/ntust-ib811-wrist-vein-dataset
• FYO [3]: Available upon request
• PUT [4]: Available upon request

Model Architecture (click images to enlarge)

• Hybrid Siamese Neural Network:

Training Workflow (click to enlarge)

🔧 GPU and CUDA Environment

• NVIDIA RTX3060 (12 GB VRAM)
• CUDA 11.2
• cuDNN 8.1.1

How to Use

Install dependencies:

pip install -r .\requirements.txt

Run training:

python .\main.py

Reference

[1] D. Romano, C. Mennella, and M. Lapegna, ‘‘Adeep learning-based method for efficient floating garbage debris recognition on high-performance edge computing platform,’’ Future Gener. Comput. Syst., vol. 174, p. 108000, 2026, doi: 10.1016/j.future.2025.108000.

[2] Sheng-Yan Dai, "NTUST-IB811 Wrist Vein Dataset", IEEE Dataport, November 13, 2025, doi:10.21227/w3ec-br30

[3] Ö. Toygar, F. O. Babalola and Y. Bitirim, ‘‘FYO: A Novel Multimodal Vein Database With Palmar, Dorsal and Wrist Biometrics,’’ IEEE Access, vol. 8, pp. 82461-82470, 2020, doi: 10.1109/ACCESS.2020.2991475.

[4] R. Kabaciński and M. Kowalski, ‘‘Vein pattern database and benchmark results,’’ Electron. Lett., vol. 47, no. 20, pp. 1127–1128, Oct. 2011, doi: 10.1049/el.2011.1441.