| dc.description.abstract | Quantum machine learning combines the realms of quantum computing with artificial intelligence, providing novel approaches to problem-solving. Quantum image recognition is one such problem that has attracted significant attention. However, many existing algorithms face a common challenge -- current quantum hardware has limited available qubits. Based on the quantum machine learning framework anointed quantum feed-forward neural network, this thesis proposes novel data encodings for image recognition. By fully leveraging the capabilities of quantum computing, these new encodings represent more of each pixel's information without requiring additional qubits. Specifically, the encodings are devised to characterise continuous pixel values, three-channel binary pixel values and three-channel continuous pixel values. In particular, the three-channel encodings enable the model to consider colours in images for increased performance. Through extensive evaluation of various datasets, the proposed encodings are demonstrated to enhance the performance of the quantum feed-forward neural network. Furthermore, the model is extended to the multi-class problem, further showcasing the improvement which the encodings provide. The work done herein is exploratory by nature, tested through classical simulations of quantum computers. As many quantum image recognition models utilise many of the same principles as the quantum feed-forward neural network, it is reasonable to imagine that the ideas and insight present in this thesis can be applied to such other models. | |
