Electronic Thesis and Dissertation Repository


Master of Engineering Science


Electrical and Computer Engineering


Jagath Samarabandu


Machine vision systems can vary greatly in size and complexity depending on the task at hand. However, the purpose of inspection, quality and reliability remains the same. This work sets out to bridge the gap between traditional machine vision and computer vision. By applying powerful computer vision techniques, we are able to achieve more robust solutions in manufacturing settings. This thesis presents a framework for applying powerful new image classification techniques used for image retrieval in the Bag of Words (BoW) framework. In addition, an exhaustive evaluation of commonly used feature pooling approaches is conducted with results showing that spatial augmentation can outperform mean and max descriptor pooling on an in-house dataset and the CalTech 3D dataset. The results for the experiments contained within, details a framework that performs classification using multiple view points. The results show that the feature encoding method known as Triangulation Embedding outperforms the Vector of Locally Aggregated Descriptors (VLAD) and the standard BoW framework with an accuracy of 99.28%. This improvement is also seen on the public Caltech 3D dataset where the improvement over VLAD and BoW was 5.64% and 12.23% respectively. This proposed multiple view classification system is also robust enough to handle the real world problem of camera failure and still classify with a high reliability. A missing camera input was simulated and showed that using the Triangulation Embedding method, the system could perform classification with a very minor reduction in accuracy at 98.89%, compared to the BoW baseline at 96.60% using the same techniques. The presented solution tackles the traditional machine vision problem of object identification and also allows for the training of a machine vision system that can be done without any expert level knowledge.