The following project outlines the results from successfully implementing convolutional and fully connected neural networks to compare which provides the best performance for the application of image classification. As well as evaluating how these networks compare with a traditional image classifier, in this case the support vector machine. Implementations were built using open source machine learning libraries, TensorFlow and Scikit-Learn. Numerous numbers of architecture designs were developed, tested and comparatively evaluated based on performance metrics. Research into literature surrounding the image classification domain indicated that a convolutional approach should outperform all other classification methods. Results from testing and comparative analysis support the findings from the literature review, helping to prove that a convolutional neural network outperforms a fully connected network and support vector machines in the context of an image classification.
PLEASE NOTE: You must be a member of the University of Lincoln to be able to view this dissertation. Please log in here.