Accelerating deep learning inference via layer truncation and transfer learning for fingerprint classification

Abstract

AbstractBiometric systems have been used extensively in the identification and verification of persons. Fingerprint biometrics stands out as the most effective due to their characteristics of Permanence, uniqueness, ergonomics, throughput, low cost, and lifelong usability. By reducing the number of comparisons, biometric recognition systems can effectively deal with large‐scale databases. Fingerprint classification is an important task used to reduce the number of comparisons by dividing fingerprints into classes. Deep learning models have demonstrated impressive classification performance in fingerprint classification tasks. The high‐level features of deep learning models can affect the transfer learning in deep learning models. Furthermore, the high‐level features involve high computational costs that can render difficulty in the deployment of the applications. This work proposes an improved system for fingerprint classification through the truncation of layers and transfer learning. Our approach modifies the ResNet50 model to improve its network inference speed and performance in fingerprint classification by removing some deep convolutional layers. We then finetune the modified model and train it using a fingerprint dataset. The results show that the finetuned modified model improves classification accuracy at a reduced computational cost. At only 5.1M parameters, our model obtained a classification accuracy of 93.3% and precision of 93.4% performing better than previous studies based on its size‐performance ratio.