Ridgeformer

Ridgeformer: Multi-Stage Contrastive Training For Fine-grained Cross-Domain Fingerprint Recognition

Accepted in IEEE International Conference on Image Processing 2025

<a href="https://github.com/KNITPhoenix/Ridgeformer"> <img src="https://github.githubassets.com/assets/GitHub-Mark-ea2971cee799.png" alt="GitHub Logo" style="width:20px; vertical-align: middle; display: inline-block; margin-right: 5px; margin-left: 5px; margin-top: 0px; margin-bottom: 0px;"/> <span class="link-text" style=" margin-right: 5px;">GitHub</span> </a> | <a href="https://arxiv.org/abs/2506.01806"> <img src="https://github.com/alpayariyak/openchat/blob/master/assets/arxiv-logomark-small-square-border.png?raw=true" alt="ArXiv Logo" style="width:20px; vertical-align: middle; display: inline-block; margin-right: 5px; margin-left: 5px; margin-top: 0px; margin-bottom: 0px;"/> <span class="link-text" style="margin-right: 5px;">Paper</span> </a> | <a href="https://paperswithcode.com/paper/ridgeformer-mutli-stage-contrastive-training"> <svg xmlns="http://www.w3.org/2000/svg" alt="ArXiv Logo" style="width:20px; vertical-align: middle; display: inline-block; margin-right: 5px; margin-left: 5px; margin-top: 0px; margin-bottom: 0px;"> <span class="link-text" style="margin-right: 5px;">PapersWithCode</span></svg> </a>

Installations and environment creation

• conda create -n ridgeformer python=3.8.19
• conda activate ridgeformer
• pip install -r requirements.txt

It requires timm version 0.5.0 and can be installed with the given .whl file. We used pytorch>=2.2.2 for CUDA=12.2

Preparing data

Datasets used in training and their application link

• Ridgebase(RB)
• The Hong Kong Polytechnic University Contactless 2D to Contact-based 2D Fingerprint Images Database Version 1.0
• IIITD SmartPhone Fingerphoto Database V1 (ISPFDv1)
• IIITD SmartPhone Finger-Selfie Database V2 (ISPFDv2)

Testing dataset:

• The Hong Kong Polytechnic University Contactless 2D to Contact-based 2D Fingerprint Images Database Version 1.0 (HKPolyU)
• Ridgebase (RB)

Preprocessing ISPFD v1 and v2 datasets

• scipts in ISPFD_preprocessing directory are used to segment out contactless images in ISPFD dataset
• requires SAM checkpoint and openai clip
• can be used after installing segment-anything and downloading SAM checkpoint
• For more information, refer to SAM's official repository Link

Manifest files creation for dataloaders

• script data_folder_creation.py in datasets directory is used to arrange the datasets in a specific folder structure

  Subject --------->
                    finger ---------->
                                      background and instances

• script manifest_file_creation.py in datasets directory is used to create manifest files used in dataloaders. The manifest file structure will be as follows:

{
Unique_finger_id_1:{
  'Contactless': ( list of paths of all contactless images )
  'Contactbased': ( list of paths of all contactbased images )
},
Unique_finger_id_2:{
  'Contactless': ( list of paths of all contactless images )
  'Contactbased': ( list of paths of all contactbased images )
}
......}

Training

Stage 1 - train_combined.py is used to train the model on Stage 1 of our architecture

Stage 2 - train_combined_fusion.py is used to train the model on Stage 2 of our architecture

All the performance ROCs and matrices are saved in combined_models_scores directory

All tensorboard logs are saved in experiment_logs directory

Testing and Evaluation

HKPolyU

• Evaluation of HKPolyU testing dataset on finetuned checkpoint from Stage 1 can be done using hkpoly_evaluation_phase1.py
• Evaluation of HKPolyU testing dataset on finetuned checkpoint from Stage 2 can be done using hkpoly_evaluation_phase2.py

Ridgebase

• Evaluation of ridgebase testing dataset on pretrained checkpoint from Stage 1 can be done using rb_evaluation_phase1.py
• Evaluation of ridgebase testing dataset on pretrained checkpoint from Stage 2 can be done using rb_evaluation_phase1.py

Performance compared with SOTA methods on HKPolyU dataset (1:1 verification)

Performance compared with SOTA methods on HKPolyU dataset (1:N identification)

Performance compared with SOTA methods on Ridgebase dataset (1:1 verification)

Performance compared with SOTA methods on Ridgebase dataset (1:N identification)

Contact

For more information or any questions, feel free to reach us at [email protected]