Bashir, Sirosh2023-04-112023-04-112022https://hdl.handle.net/10539/34954A dissertation submitted in fulfillment of the requirements for the degree of Master of Science in Engineering to the Faculty of Engineering and the Built Environment, School of Electrical and Information Engineering, University of the Witwatersrand, Johannesburg,2022Face detection and recognition in real-time video with low illumination have become important for various applications such as smart cities, airports, underground mines (access control, rescue operations, etc.). However, most of the legacy computer vision techniques are not been able to detect and recognize faces from such images with acceptable accuracy. There is, therefore, a need to explore machine learning techniques to achieve face detection and recognition in low illumination images with acceptable accuracy. Face detection, registration, and recognition in unconstrained environments have been popular research fields and this has resulted in numerous new methodologies for face detection and recognition in the recent past. One of the most popular techniques in this regard which performs well with good illumination images are the method based on Convolutional Neural Network with standard configurations. Although ample amount of work has been done but low illumination research was not well addressed in the literature review[1][2]. Also, most of the CNN based techniques provide promising results for still images[3]. Thus, there is a need to understand the CNN configuration for low illumination videos. The research aims to explore the potential of different machine learning approaches for face detection and recognition in a low illumination environment. Different machine learning approaches such as YOLO and LBPH with KNN were used to analyze the accuracy of detection and recognition in a low illumination environment. A number of techniques were applied to improve the detection and recognition accuracy in low illumination. For example, data augmentation was implemented to increase the diversity of the dataset which was later on used for training purposes. Similarly, preprocessing techniques were applied to a standard dataset to improve the accuracy of the system. The exploratory data analysis technique runs on a preprocessed annotated images dataset. Exploratory data analysis (EDA) gives insight from the dataset, which assists in hyperparameter tuning for training and testing purpose[4]. The simulation results shows that YOLO outperforms with LBPH with KNN for face detection and recognition in a low illumination environment.enDissertation