Latent Fingerprints Classification Using Transfer Learning

Transfer learning refers to reusing the learned weights of a trained model into another untrained target model instead of training the target model from scratch. Transfer learning is the next driver of Machine Learning success after supervised learning. In this paper, transfer learning is used for classification of latent fingerprints. A pre-trained Convolutional Neural Network (CNN) model (AlexNet) is fine-tuned to classify latent fingerprints into four different classes. The pre-trained model is fine-tuned on NIST-SD4 plain fingerprint and IIIT-D latent databases. By using a pre-trained model, the target model retains general features from the base model and learns specific features by fine tuning it on target dataset. The technique has been demonstrated on latent fingerprints from IIIT-D latent database and achieved the classification accuracy of around 80%.