DEVELOPING AND EVALUATING A HYBRID QUANTUM CONVOLUTIONAL NEURAL NETWORK FOR ENHANCED IMAGE CLASSIFICATION

Abstract

In image classification problems, the size of the space and parameter increase are limiting factors to implement a classical convolutional neural network (CNN) on resource-constrained systems. This paper investigates hybrid quantum-classical learning models to use CNN as a space-saving alternative. Two hybrid quantum neural network architectures are designed, namely HQNN-Parallel and HQNN-Quanv. HQNN-Parallel is one of such quantum neural networks: it is a hybrid network consisting of a classical convolutional feature extractor and parallel four-qubit quantum circuits in the hybrid, amplitude encoding, trainable rotation gate, CNOT-based entanglement, and Pauli-Z measurement. HQNN-Quanv proposes to use a quanvolutional layer to help in feature transformation and image-resolution reduction. Tested on MNIST, Medical MNIST, and CIFAR-10 datasets and comes with metrics including accuracy, loss, precision, recall, F1 score, and trainable parameters. HQNN-Parallel on MNIST is much more accurate (99.21%) than the classical CNN baseline, which is approximately 100,000 parameters. The accuracy of the hybrid model is 84.11 per cent for CIFAR-10, compared to the classical CNN baseline accuracy of 83.12 per cent. The accuracy of the HQNN-Quanv is 99.0 % on MNIST, while the classical CNN has 99.1 % accuracy, and there is a reduction of 74.59 % of parameters in the classical CNN compared to the HQNN-Quanv. The results suggest that competitive accuracy can be obtained with hybrid quantum-classical models, with enhanced parameter efficiency. However, the research is limited in that it is only simulation-based, validating with real devices for noise and no noise. Additional research will be conducted via noise-model experiments, statistical testing (repeated experiments), and hardware construction to see if they provide an actual advantage in the real world.