Fingerprint Recognition with Siamese Network

Given a fingerprint image, model identifies a match from a database of fingerprints.

Procedure

• Training Fingerprint Model

  Siamese network is trained on a datastore of fingerprint images from different people. Each fingerprint is a set of 10 prints from each finger

  Fingerprints used during training are stored for future matching. To reduce computational time, we store only the extracted (intermediate) features.

  Model is trained, validated and stored.

  Model training code is in code/train.py .

  To run model training

  from train import ModelBuilder
  
  imagedir = '..{0}data{0}images'.format(os.sep)
  modelBuilder = ModelBuilder(imagedir)
  modelBilder.trainModel()

• Fingerprint Match After model has been trained, fingerprint matching modelfile is saved in model/siamese_model.h5

  To run a fingerprint match against a set of templates (stored during training)

  from match import FingerprintAuthentication
  
  matchthreshold = 0.8
  fingerprint_to_match = '..{0}data{0}images{0}101_4.tif'.format(os.sep)
  fingerprintAuthentication = FingerprintAuthentication()
  matched_tiffile, matched_prob, matched_personid = fingerprintAuthentication.matchFingerprint(fingerprint_to_match)
  print('Matched TIF: {}, Probability: {}, PersonID: {}'.format(matched_tiffile, matched_prob, matched_personid ))
  if matched_prob >= matchthreshold:
      print('Got a fingerprint match with {} for person id {}'.format(matched_tiffile, matched_personid))
  else:
      print('No match found')

Model Architecture

Model is a siamese neural network with pre-trained ResNet50 as feature extractor

Training and Validation Loss

Tensorboard - Training and Validation cross entropy loss

Sample Command Output

After the training completes, the validation metrics is outputed to stdout

Result of fingerprint match