A biometric system with a database of registered users and two main functions:

• USER VERIFICATION: given an image of an iris and user's ID, the system verifies whether this image matches known user patterns,
• USER IDENTIFICATION: given an image, the system predicts a possible user from the database.

The system was build on a deep neural network using transfer learning. Given a normalized picture of an iris, the model classifies it to one of the system's users, and gives a probability of the prediction. In order to reject uncertain classifications, only results with probability greater than 98% are considered to be successful.

The architecture chosen was ResNet50, which was pre-trained on ImageNet. The model was the fine-tuned to the provided database of irides pictures.

The training was done in a Google Colaboratory notebook with GPU support.

• number of classes: 50
• learning rate: 0.0002
• optimizer: Adam
• number of epochs: 100
• batch size: 16

• maximal training accuracy: 96.14%
• maximal validation accuracy: 93.00%

The system recognizes a user if the neural network underneath classifies an image with at least 98% confidence.

This project was written in Python 3.8. To create Conda environment with all the system's requirements, run:

conda env create -f conda_environment.yml

Then proceed to activate the environment before using the biometric system:

conda activate iris

This system was trained on pictures from UBIRIS.v1 database.

From this dataset, 50 people were selected to create the biometric system. Their pictures are stored at data/system_database/registered_users.

Additional 50 different people were selected to a validation set. Their pictures are stored at data/system_database/unknown_users.

The entry point to the system is through the main file biometric_system.py.

usage: biometric_system.py [-h] [-u USER] [-m MODEL] image {identify,verify}

Biometric system.

positional arguments:
  image                 Path to the image.
  {identify,verify}     Program mode. If you want to identify a user based on an image, choose 'identify'; If you want to verify whether an image portraits a particular user,
                        choose 'verify' and provide the user's ID in the next argument.

optional arguments:
  -h, --help            show this help message and exit
  -u USER, --user USER  User's ID. Only used with mode 'verify'.
  -m MODEL, --model MODEL
                        Path to the trained classifier model.

The system correctly identified a user:

python3 biometric_system.py data/system_database/registered_users/7_0.jpg identify

This image portraits user 7 (Prediction probability: 99.87%)
Program exited with code: 0 - Successfully identified a user

The system was not able to recognize a user from their image because of low prediction probability (the eye was closed on the picture):

python3 biometric_system.py data/system_database/registered_users/45_8.jpg identify

Program exited with code: 1 - Could not identify a user - user was not found in the database

Successful verification:

python3 biometric_system.py data/system_database/registered_users/61_8.jpg verify --user 61

Successfully verified user 61 (Prediction probability: 99.94%)
Program exited with code: 0 - Successfully verified a user

User was not verified, since the provided ID was different from the classification:

python3 biometric_system.py data/system_database/registered_users/61_8.jpg verify --user 85

Program exited with code: 1 - Failed to verify a user - user ID did not match the classification

The system was not able to recognize a user from their image because of low prediction probability:

python3 biometric_system.py data/system_database/registered_users/46_2.jpg verify --user 46

Program exited with code: 1 - Failed to verify a user - user was not found in the database

Submitting an image of a user that is not registered in the system:

python3 biometric_system.py data/system_database/unknown_users/1_0.jpg identify

Program exited with code: 1 - Failed to verify a user - user was not found in the database