High-Resolution Image Generation using Quantum Generative Adversarial Networks and Quantum Autoencoders

Abstract

This thesis presents the Quantum Latent Generative Adversarial Network (QL GAN) framework to address the scalability issue of quantum GAN-based image gen eration methods. The QLGAN framework comprises two quantum machine learning models: a quantum autoencoder that compresses the quantum state into fewer qubits and a quantum GAN that learns the distribution of compressed quantum states. Ad ditionally, this thesis proposes the Quantum Image Autoencoder (QIAE), which is optimized for quantum probability image encoding. A bidirectional loss function is introduced to address the mismatch between loss and quality in image reconstruction. Experimental results shows that QIAE achieves state-of-the-art performance on image compression and reconstruction tasks over MNIST and FMNIST datasets. Further more, the combination of QLGAN and QIAE generates high-quality, high-resolution quantum images without the need for classical post-processing.