Alpha matting is an important topic in areas of computer vision. It has various applications, such as virtual reality, digital image and video editing, and image synthesis. The conventional approaches for alpha matting perform unsatisfactorily when they encounter complicated background and foreground. It is also difficult for them to extract alpha matte accurately when the foreground objects are transparent, semi-transparent, perforated or hairy. Fortunately, the rapid development of deep learning techniques brings new possibilities for solving alpha matting problems. In this thesis, we propose a residual convolutional grid network for alpha matting, which is based on the convolutional neural networks (CNNs) and can learn the alpha matte directly from the original image and its trimap. Our grid network consists of horizontal residual convolutional computation blocks and vertical upsampling/downsampling convolutional computation blocks. By choosing different paths to pass information by itself, our network can not only retain the rich details of the image but also extract high-level abstract semantic information of the image. The experimental results demonstrate that our method can solve the matting problems that plague conventional matting methods for decades and outperform all the other state-of-the-art matting methods in quality and visual evaluation. The only matting method performs a little better than ours is the current best matting method. However, that matting method requires three times amount of trainable parameters compared with ours. Hence, our matting method is the best considering the computation complexity, memory usage, and matting performance.

Ulcerative colitis (UC) is a chronic inflammatory disease characterized by periods of relapses and remissions affecting more than 500,000 people in the United States. To achieve the therapeutic goals of UC, which are to first induce and then maintain disease remission, doctors need to evaluate the severity of UC of a patient. However, it is very difficult to evaluate the severity of UC objectively because of non-uniform nature of symptoms and large variations in their patterns. To address this, in our previous works, we developed two different approaches in which one is using the image textures, and the other is using CNN (convolutional neural network) to measure and classify objectively the severity of UC presented in optical colonoscopy video frames. But, we found that the image texture based approach could not handle larger number of variations in their patterns, and the CNN based approach could not achieve very high accuracy. In this paper, we improve our CNN based approach in two ways to provide better accuracy for the classification. We add more thorough and essential preprocessing, and generate more classes to accommodate large variations in their patterns. The experimental results show that the proposed preprocessing can improve the overall accuracy of evaluating the severity of UC.

Indiana University-Purdue University Indianapolis (IUPUI) / Meningioma represent more than one-third of all primary central nervous system (CNS) tumors, and it can be classified into three grades according to WHO (World Health Organization) in terms of clinical aggressiveness and risk of recurrence. A key component of meningioma grades is the mitotic count, which is defined as quantifying the number of cells in the process of dividing (i.e., undergoing mitosis) at a specific point in time. Currently, mitosis counting is done manually by a pathologist looking at 10 consecutive high-power fields (HPF) on a glass slide under a microscope, which is an extremely laborious and time-consuming process. The goal of this thesis is to investigate the use of computerized methods to automate the detection of mitotic nuclei with limited labeled data. We built computational methods to detect and quantify the histological features of mitotic cells on a whole slides image which mimic the exact process of pathologist workflow. Since we do not have enough training data from meningioma slide, we learned the mitotic cell features through public available breast cancer datasets, and predicted on meingioma slide for accuracy. We use either handcrafted features that capture certain morphological, statistical, or textural attributes of mitoses or features learned with convolutional neural networks (CNN). Hand crafted features are inspired by the domain knowledge, while the data-driven VGG16 models tend to be domain agnostic and attempt to learn additional feature bases that cannot be represented through any of the handcrafted features. Our work on detection of mitotic cells shows 100% recall , 9% precision and 0.17 F1 score. The detection using VGG16 performs with 71% recall, 73% precision, and 0.77 F1 score. Finally, this research of automated image analysis could drastically increase diagnostic efficiency and reduce inter-observer variability and errors in pathology diagnosis, which would allow fewer pathologists to serve more patients while maintaining diagnostic accuracy and precision. And all these methodologies will increasingly transform practice of pathology, allowing it to mature toward a quantitative science.

For input data of homogenous type, the standard form of convolutional neural network is normally constructed with universally applied filters to identify global patterns. However, for certain datasets, there are identifiable trends and patterns within subgroups of input data. This research proposes a convolutional neural network that deliberately partitions input data into groups to be processed with unique sets of convolutional layers, thus identifying the underlying features of individual data groups. Training and testing data are built from historical prices of stock market and preprocessed so that the generated datasets are suitable for both standard and the proposed convolutional neural network. The author of this research also developed a software framework that can construct neural networks to perform necessary testing. The calculation logic was implemented using parallel programming and executed on a Nvidia graphic processing unit, thus allowing tests to be executed without expensive hardware. Tests were executed for 134 sets of datasets to benchmark the performance between standard and the proposed convolutional neural network. Test results show that the partitioned convolution method is capable of performance that rivals its standard counterpart. Further analysis indicates that more sophisticated method of building datasets, larger sets of training data, or more training epochs can further improve the performance of the partitioned neural network. For suitable datasets, the proposed method could be a viable replacement or supplement to the standard convolutional neural network structure. / Thesis / Master of Applied Science (MASc) / A convolutional neural network is a machine learning tool that allows complex patterns in datasets to be identified and modelled. For datasets with input that consists of the same type of data, a convolutional neural network is often architected to identify global patterns. This research explores the viability of partitioning input data into groups and processing them with separate convolutional layers so unique patterns associated with individual subgroups of input data can be identified. The author of this research built suitable test datasets and developed a (parallel computation enabled) framework that can construct both standard and proposed convolutional neural networks. The test results show that the proposed structure is capable of performance that matches its standard counterpart. Further analysis indicates that there are potential methods to further improve the performance of partitioned convolution, making it a viable replacement or supplement to standard convolution.

Direction of Arrival (DOA) estimation with digital arrays under unknown Gaussian distributed element location perturbation has detrimental effects to the performance of traditional DOA estimation techniques. This work proposes an artificial intelligence (AI) approach as a solution to this problem. A Deep Convolutional Neural Network (DCNN) is proposed and experimentation into network parameters, classification networks, and how the DCNN is applied to the DOA problem are studied. It is shown that this AI based approach is successful in estimating the DOA with perturbed arrays where traditional approaches fail.

Los ciberdelincuentes están en constante desarrollo de nuevas técnicas para evadir las medidas de seguridad implementadas por expertos e investigadores, lo que permite que el malware evolucione rápidamente. Además, detectar malware en varios sistemas es un desafío debido a que cada entorno informático tiene sus propias características únicas. Las técnicas tradicionales, como la detección de malware basada en firmas, han perdido efectividad y han sido en gran medida reemplazadas por enfoques más modernos, que incluyen el ML y la detección de amenazas basada en el comportamiento robusto en múltiples plataformas. Los investigadores aplican estas técnicas a diversas fuentes de datos, como el tráfico de red, binarios y datos de comportamiento, para extraer características relevantes y alimentar modelos que permiten hacer predicciones precisas. El objetivo de esta investigación es proporcionar un nuevo conjunto de datos compuesto por una cantidad sustancial de muestras de alta calidad basadas en el comportamiento del software. Debido a la falta de un formato de representación estándar para el comportamiento del malware en la investigación actual, también presentamos un método innovador para representar el comportamiento del malware, convirtiendo las llamadas a API en imágenes 2D. Además, proponemos y describimos la implementación de un nuevo modelo de ML basado en la clasificación binaria (malware o software benigno) utilizando el conjunto de datos novedoso mencionado anteriormente, lo que establece una base para su evaluación. Hemos realizado experimentos exhaustivos, validando el modelo propuesto tanto con el conjunto de datos propuesto como con datos del mundo real. En términos de métricas, este supera a un modelo conocido que también se basa en el análisis del comportamiento y tiene una arquitectura similar.

Face recognition is the problem of identifying individuals in images. This thesis evaluates two methods used to determine if pairs of face images belong to the same individual or not. The first method is a combination of principal component analysis and a neural network and the second method is based on state-of-the-art convolutional neural networks. They are trained and evaluated using two different data sets. The first set contains many images with large variations in, for example, illumination and facial expression. The second consists of fewer images with small variations. Principal component analysis allowed the use of smaller networks. The largest network has 1.7 million parameters compared to the 7 million used in the convolutional network. The use of smaller networks lowered the training time and evaluation time significantly. Principal component analysis proved to be well suited for the data set with small variations outperforming the convolutional network which need larger data sets to avoid overfitting. The reduction in data dimensionality, however, led to difficulties classifying the data set with large variations. The generous amount of images in this set allowed the convolutional method to reach higher accuracies than the principal component method.

Object recognition systems have significant influences on modern life. Face, iris and finger point recognition applications are commonly applied for the security purposes; ASR (Automatic Speech Recognition) is commonly implemented on speech subtitle generation for various videos and audios, such as YouTube; HWR (Handwriting Recognition) systems are essential on the post office for cheque and postcode detection; ADAS (Advanced Driver Assistance System) are well applied to improve drivers’, passages’ and pedestrians’ safety. Object recognition techniques are crucial and valuable for academia, commerce and industry. Accuracy and efficiency are two important standards to evaluate the performance of recognition techniques. Accuracy includes how many objects can be indicated in real scene and how many of them can be correctly classified. Efficiency means speed for system training and sample testing. Traditional object detecting methods, such as HOG (Histogram of orientated Gradient) feature detector combining with SVM (Support Vector Machine) classifier, cannot compete with frameworks of neural networks in both efficiency and accuracy. Since neural network has better performance and potential for improvement, it is worth to gain insight into this field to design more advanced recognition systems. In this thesis, we list and analyze sophisticated techniques and frameworks for object recognition. To understand the mathematical theory for network design, state-of-the-art networks in ILSVRC (ImageNET Large Scale Visual Recognition Challenge) are studied. Based on analysis and the concept of edge detectors, a simple CNN (Convolutional Neural Network) structure is designed as a trail to explore the possibility to utilize the network of high width and low depth for region proposal selection, object recognition and target region refining. We adopt Le-Net as the template, taking advantage of multi-kernels of GoogLe-Net. We made experiments to test the performance of this simple structure of the vehicle and face through ImageNet dataset. The accuracy for the single object detection is 81% in average and for plural object detection is 73.5%. We refined networks through many aspects to reach the final accuracy 95% for single object detection and 89% for plural object detection.

Multiphase flow provides a solution to the high heat flux and precision required by modern-day gadgets and heat transfer devices as phase change processes make high heat transfer rates achievable at moderate temperature differences. An application of multiphase flow commonly used in industry is the condensation of refrigerants in inclined tubes. The identification of two-phase flow patterns, or flow regimes, is fundamental to the successful design and subsequent optimisation given that the heat transfer efficiency and pressure gradient are dependent on the flow structure of the working fluid. This study showed that with visualisation data and artificial neural networks (ANN), a machine could learn, and subsequently classify the separate flow patterns of condensation of R-134a refrigerant in inclined smooth tubes with more than 98% accuracy. The study considered 10 classes of flow pattern images acquired from previous experimental works that cover a wide range of flow conditions and the full range of tube inclination angles. Two types of classifiers were considered, namely multilayer perceptron (MLP) and convolutional neural networks (CNN). Although not the focus of this study, the use of a principal component analysis (PCA) allowed feature dimensionality reduction, dataset visualisation, and decreased associated computational cost when used together with multilayer perceptron neural networks. The superior two-dimensional spatial learning capability of convolutional neural networks allowed improved image classification and generalisation performance across all 10 flow pattern classes. In both cases, the classification was done sufficiently fast to enable real-time implementation in two-phase flow systems. The analysis sequence led to the development of a predictive tool for the classification of multiphase flow patterns in inclined tubes, with the goal that the features learnt through visualisation would apply to a broad range of flow conditions, fluids, tube geometries and orientations, and would even generalise well to identify adiabatic and boiling two-phase flow patterns. The method was validated by the prediction of flow pattern images found in the existing literature. / Dissertation (MEng)--University of Pretoria, 2021. / NRF / Mechanical and Aeronautical Engineering / MEng / Restricted

The aim of this thesis is to study problems of deep convolutional neural networks and the connected classification of images and to experiment with the architecture of particular network with the aim to get the most accurate results on the selected dataset. The thesis is divided into two parts, the first part theoretically outlines the properties and structure of neural networks and briefly introduces selected networks. The second part deals with experiments with this network, such as the impact of data augmentation, batch size and the impact of dropout layers on the accuracy of the network. Subsequently, all results are compared and discussed with the best result achieved an accuracy of 86, 44% on test data.