IMPROVING THE PERFORMANCE OF HUMAN EAR-BASED AUTOMATED BIOMETRIC SYSTEMS USING FEATURES FUSION AND COMPARING WITH CONVOLUTION NEURAL NETWORKS

Abstract

In this study, the features fusion method was used to improve the performance of human ear-based automated biometric system recognizing and identifying individuals without the need for user’s interaction. The image ray transform (IRT) algorithm was used in the recognition stage. Subsequently, binarised statistical image features (BSIF) and weber local descriptor (WLD) were extracted from the normalized outputs of the recognition stage. The canonical correlation analysis (CCA) method was then used to combine the two extracted feature vectors and create a distinct feature vector. Finally, the K-nearest neighbors’ algorithm (k-NN) was used for decision making. The accuracy of designing and implementing this system was 99.82%. Then a convolution neural network (CNN) was designed and trained with binarised statistical image features (BSIF) and weber local descriptor (WLD). The simulation accuracy of CNN was 100%. According to the results of this study, convolution neural network based on BSIF and WLD features has superior performance over state-of-the-art methods.