ITSEC - Reconstruction of Fingerprints from Minutiae Templates

Contributors (Project Team)

• Venkata Srinath Mannam
• Meghana Rao
• Lukas Partes

Getting Started

• In this work we have explored the pix2pix - image to image translation paper for generating realistic fingerprints from minutiae maps.
• This repository particularly is used for generating the training data(In other words pre-processing) for the pix2pix model.
• These instructions will get you to set up on your local machine for development and testing purposes.

Prerequisites

• Should have fingerprints dataset.
• Should have access to the softwares like Verifinger or NIST
• Understanding of python, linux would be better.
• Create a conda env or python venv

Installing

• Clone the repository or download and unzip it.
• Make sure that all the files are present in folder and in the following similar structure.

  ITSEC21(Parent Folder)
      evaluation
              evaluation_normal_multi_folders.py
              evaluation_normal_single_folder.py
      minutiae_map
      tools
      utils
      .gitignore
      README.md
  

• Install the packages mentioned in environment.yml

   #Do this in the project folder console.
   conda env create -f environment.yml
  

Working in devlopment

The project includes multiple tasks:

• Getting Datasets
• Extrcating Minutiae
• Minutiae Map Construction
• Training Data Creation
• Data Augmentation
• Evaluation

Getting Datasets

• We have used two public datasets in this project Cross Match and U.are.U

Extracting Minutiae

• To extract minutiae we have used MINDTCT tool by NIST.
• The mentioned datasets are in ".tif" format but MINDTCT tool expects the inputs to be in ".jpeg" format. To convert the datasets, we have used the python script tools/convert_TIF_JPEG.py
• To iteratively extract minutiae we have used the script file src/minutiae_Map_Reconstruction/extractMinutiae.sh

Minutiae Map Construction

• We have taken .xyt files which contains minutiae information.
• Run the python file of minutiae_map/main.py.
• May be input folder and target folder paths needs to be modified in this file.

Training Data Creation

• First for each dataset we create the images in the following way:

  • Cross Match dataset (Both for fingerprints, minutiae maps)
    • the original dimensions are 504*480.
    • we have padded them, so that the shape is 504*504.
    • Now we downscale the images until 256*256.
    • The above 3 steps are performed by using the python script tools/downsample_imgs_to_256.py
  • U.are.U dataset (Both for fingerprints, minutiae maps)
    • the original dimensions are 326*357.
    • we have padded them, so that the shape is 357*357.
    • Now we downscale the images until 256*256.
    • The above 3 steps are performed by using the python script tools/downsample_imgs_to_256.py

• We should create the train, test, val sets for both minutiae-maps and fingerprints(from downsampled images). This can be done by running the python script tools/train_test_val_split.py

• Using the above created train, test, val sets and follow the instructions of pix2pix-network for training, testing the model.

Data Augmentation

• In order to perform data augmentation for the dataset as the size of the datasets are very low, run the python script tools/data_augmentation.py
• In our research we found that data-augmentation doesn't help the model training. Instead, we should increase the dataset size.

Evaluation

• Once the model is trained, now run the model for testing and it will give us the reconstructed fingerprints.
• Now, we should perform the evaluation for the original fingerprints and reconstructed fingerprints generated by the model.
• The evaluation can be done by running the python script evaluation/evaluation_normal_single_folder.py.
• The above step will work only if you have configured Verifinger evaluation in your system.

Note

• The mentioned Folder-Paths needs to be corrected in the mentioned python script files.
• Please ignore the other files as we have tried with some other datasets(for example casia) and additional experiments, so we had written new scripts.