Abstract
In modern society, there is an ever-growing need to determine the identity of a person in many applications including computer security, financial transactions, borders, and forensics. Early automated methods of authentication relied mostly on possessions and knowledge. Notably these authentication methods such as passwords and access cards are based on properties that can be lost, stolen, forgotten, or disclosed. Fortunately, biometric recognition provides an elegant solution to these shortcomings by identifying a person based on their physiological or behaviourial characteristics. However, due to the diverse nature of biometric applications (e.g., unlocking a mobile phone to cross an international border), no biometric trait is likely to be ideal and satisfy the criteria for all applications. Therefore, it is necessary to investigate novel biometric modalities to establish the identity of individuals on occasions where techniques such as fingerprint or face recognition are unavailable. One such modality that has gained much attention in recent years which originates from forensic practices is the lip. This research study considers the use of computer vision methods to recognise different lip prints for achieving the task of identification. To determine whether the research problem of the study is valid, a literature review is conducted which helps identify the problem areas and the different computer vision methods that can be used for achieving lip print recognition. Accordingly, the study builds on these areas and proposes lip print identification experiments with varying models which identifies individuals solely based on their lip prints and provides guidelines for the implementation of the proposed system. Ultimately, the experiments encapsulate the broad categories of methods for achieving lip print identification. The implemented computer vision pipelines contain different stages including data augmentation, lip detection, pre-processing, feature extraction, feature representation and classification. Three pipelines were implemented from the proposed model which include a traditional machine learning pipeline, a deep learning-based pipeline and a deep hybridlearning based pipeline. Different metrics reported in literature are used to assess the performance of the prototype such as IoU, mAP, accuracy, precision, recall, F1 score, EER, ROC curve, PR curve, accuracy and loss curves. The first pipeline of the current study is a classical pipeline which employs a facial landmark detector (One Millisecond Face Alignment algorithm) to detect the lip, SURF for feature extraction, BoVW for feature representation and an SVM or K-NN classifier. The second pipeline makes use of the facial landmark detector and a VGG16 or ResNet50 architecture. The findings reveal that the ResNet50 is the best performing method for lip print identification for the current study. The third pipeline also employs the facial landmark detector, the ResNet50 architecture for feature extraction with an SVM classifier. The development of the experiments is validated and benchmarked to determine the extent or performance at which it can achieve lip print identification. The results of the benchmark for the prototype, indicate that the study accomplishes the objective of identifying individuals based on their lip prints using computer vision methods. The results also determine that the use of deep learning architectures such as ResNet50 yield promising results.
M.Sc. (Science)