Fingerprint pattern classification using deep transfer learning and data augmentation

Decreasing the number of matching comparisons between presented fingerprints and their respective templates in automated fingerprint identification systems (AFIS) is salient, especially when dealing with large databases. Fingerprint classification abets the achievement of this goal by stratifying fingerprints into their respective pattern profiles. However, the significant inter-class variation among patterns and minor intra-class variations among fingerprint patterns belonging to a similar class remains a obstacle. Unlike the verification process of fingerprints that requires 1:1 matching of templates, the identification process of fingerprint patterns requires 1:N matching to attest the presence of fingerprint in the database, which leads to a higher number of comparisons. Motivated by this problem, we employed the use of deep transfer learning and data augmentation to develop a fingerprint pattren clasifier to clasify six fingerprint patterns. Three separate models were birth from the utilization of the VGG16, VGG19, and DenseNet121 pre-trained models following some preliminary experiment. Results from the implementation of the proposed deep transfer learning with some data augmentation schemes on the selected VGG16, VGG19, and DenseNet121 pre-trained models manifested classification accuracy of 98.2%, 97%, and 97.8%, respectively, as compared to the 93.9%, 93.7% and 92% rendered by the same models devoid of data augmentation. Hence, experimental results proved that data augmentation improves the efficacy of fingerprint pattern classifier models.