The Use of Neural Networks for Identity Recognition Through Palm Print and Voting Technique

Abstract

Biometry has emerged as the best solution for criminal identification and access control applications where resources or information need to be protected from unauthorized access. Biometric traits such as fingerprint, face, palm print, iris, and handgeometry have been well explored; and matured approaches are available to perform personal identification. The human hand is one of the body parts with special characteristics that are unique to every individual. The distinctive features can give some information about an individual, thus, making it a suitable body part that can be relied upon for biometric identification and, specifically, gender recognition. Several studies have suggested that the hand has unique traits that help in gender classification. Human hands form part of soft biometrics as they have distinctive features that can give information about a person. Nevertheless, the information retrieved from the soft biometrics can be used to identify an individual’s gender. Furthermore, soft biometrics can be combined with the main biometrics characteristics that can improve the quality of biometric detection. Gender classification using hand features such as palm contributes significantly to the biometric identification domain and, hence, presents itself as a valuable research topic. Despite a period of remarkable evolution, no extensive comparison and evaluation have been performed up till now to study the effect of the representation of data through the descriptors on palmprint recognition problem. Motivated by this statement, this research aims to fill this gap and provide a comprehensive comparative study of the performance of a large number of recent state-of-the-art texture descriptors in palmprint recognition. The research emphasizes the opportunities for features representation and analysis from a palmprint image using handcrafted (Curvelet, Wavelet, Wave Atom, SIFT, Gabor, LBP) and neural approaches based convolutional neural network. All previous features were merged at the decision level by a proposed voting method in order to enhance the identification of a person. The proposed approach was tested in a number of experiments on the CASIA, IITD, and 11k palmprint databases. The testing yielded positive results supporting the use of the described voting technique for human recognition purposes. This research also explores the use of Discrete Wavelet Transform (DWT) in gender identification, with SqueezeNet acting as a tool for unsheathing features, and Support Vector Machine (SVM) operating as a discriminative classifier. In this research, we aim also to apply fusion at score level on predictive labels obtained from different descriptors rather than the labels obtained from different classifiers. We study also the effect of using fusion at decision level through Mode Voting Technique (MVT) to achieve a good performance of our proposed system for identity recognition. From the results, it is clear that the fusion at decision level using the Mode Voting Technique guarantees an excellent recognition rate regardless of low recognition rate of some datasets. The mode voting technique ranks top of the list of SVM classifiers used for each database.