Image classification via transfer learning

Image classification has been used in many real-world applications such as self-driving cars, recommender systems and recognition systems. With the advent of deep learning and artificial intelligence, the accuracy of such systems has increased greatly. However, such methods require large amounts of labelled data during the training phase in order to achieve a high level of accuracy. This data may not be available in all cases. Furthermore, domain shifts which occur in many real-world situations may also affect the accuracy of the classifier. This project studies the use of transfer learning, in particular, domain adaptation on Convolutional Neural Networks (CNNs) to improve classifier accuracy in cases of domain shifts where labelled data is not available. Firstly, the effectiveness of different domain adaptation algorithms is explored on well-known online digit datasets as well as a custom Dark-Images dataset (Chapter 5). Secondly, a graphical user interface (GUI) is developed to compare the performance of domain adaptation methods on an image-by-image basis (Chapter 6). Thirdly, class alignment during domain adaptation training is explored (Chapter 7); a novel framework is proposed that shows promising performance improvements to existing domain adaptation algorithms (Chapter 8). Lastly, this report also gives recommendations for implementing domain adaptation and on future works to explore (Chapter 9).Bachelor of Engineering (Electrical and Electronic Engineering