Abstract:

Abstract A biometric recognition is performed with feature extraction, matching, and classification. Before the emergence of deep learning, biometric recognition has completely relied on manual feature extraction. Convolutional neural networks have automated feature extraction. To fetch features manually, domain knowledge and programming expertise are required. A dataset quality affects accuracy of a shallow classifier whereas the performance of a deep learning model succeeds in providing high accuracy only if the training dataset is balanced, qualitative, and large enough to distinguish features from various classes. Constructing a classifier from a large training dataset is time-consuming and causes overfitting. On the other hand, a small dataset-based model suffers from underfitting. To overcome the said issues, this paper proposed a hybrid approach of a concatenation of manually extracted domain-independent features such as Kekre’s Median Codebook and Kekre’s Fast Codebook and automatically extracted features through CNNs by processing samples from physiological and behavioral biometric traits independently and feeding these to neural networks to achieve best possible accuracy of classification so that the possibility of underfitting and overfitting is avoided. This method is evaluated by applying it to LFW, UPOL, IITD, IITD V1, and UserSignatureDatabase datasets of face, iris, fingerprint, palmprint, and signature respectively, and resulting models achieved improved (certain models achieved equivalent accuracy) with reduced memory and learning time.

Keywords: