Current Students¶

The following are postgraduate students currently under my supervision or co-supervision.

Project Students¶

Computer Science Honours¶

Student	Project Title	Abstract
Kyle Johnston 21619166@sun.ac.za	Face and Occlusion Recognition using Pairwise Differential Siamese Networks	The presence of occlusions has long been a challenge in the field of facial recognition. Inspired by the human’s ability to focus on unobstructed facial features, the goal of this project is to develop a facial recognition system with the ability to recognize occlusions and remove their effects from the facial recognition process. Using Siamese networks, a mask learning strategy that caters for commonly found occlusions such as masks and glasses is proposed. This project explores the use of a segmentation network along with the mask learning strategy and existing state-of-the-art convolutional neural network architectures for image recognition to tackle the challenge of occlusions in facial recognition.
Thomas Scholtz 21681147@sun.ac.za	A Robust Approach To Real World Anomaly Detection	The task of anomaly detection in surveillance footage spans a variety of contrasting situations, each with differing versions of normal and abnormal activity. The objective of this research is to develop a framework which detects anomalies in CCTV surveillance, irrespective of the context of the footage i.e., provided with a sufficient quantity of normal footage, the framework is to distinguish what is abnormal in that scenario.The implementation takes a weakly-supervised learning approach which defines anomaly detection as a regression problem. The framework will be trained and tested on the UCF-Crime dataset which provides a realistic replication of the task and therefore a worthy challenge.Keywords: Deep Learning, 3D Convolutional Networks, Artificial Neural Networks, Multiple Instance Learning, Appearance-Motion Correspondence.
Dionne T Chasi 20854277@sun.ac.za	Multi Fram Quality ENhancement using Quantized Data	Video enhancement paves the way for other automated video processes such as detection, recognition,analysis, segmentation, and surveillance. Over the past few years deep learning has proven to provide improved results for video enhancement. Recent approaches that have used deep learning for video enhancement do not take advantage of the neighboring frames when processing a single frame. The aim of this project is to extend and improve the Multi-Frame Quality Enhancement (MFQE) approach. The MFQE approach consists of Peak Quality Frame (PQF) Detection using a Bi LSTM (Bi directional Long Short Memory) based detector and video frame enhancement using a Multiframe Convolutional Network (MFCNN). In this research we will make use of quantized data to improve the accuracy and computation time of the MFQE. The quantized data will be obtained from parsing data into a deep belief network which has a SoftMax classifier as the outer layer. The output from the SoftMax classifier will be considered as the quantized data. The Quantized data will be fed into the MFCNN and the PQF’s are used to improve the non-PQF's.
MJ Adendorff mj.praetor@gmail.com	Deep learning for leafroll disease detection	Leafroll disease affects many of the grape farms in South Africa and is typically detected too late.This project aims to research the possibility of creating accurate deep learning models using a small dataset of locally collected images. The research is divided into two main stages, the first of these is concerned with the use of two known methods for Few-shot learning, namely Prototypical networks and Siamese neural networks, testing them both with and without meta learning. The second questions the possibility of generating synthetic images to form a dataset which can be used to train a deep learning model in a conventional way. The goal is to ultimately have a functional deep learning model that can accurately classify the Leafroll disease in real-case scenarios.

Masters Students¶

Computer Science¶

Student	Thesis Title	Abstract