Image Generation And Classification Algorithms Based On Quantum Machine Learning

Machine learning has become a hot field of artificial intelligence in recent decades,and image generation and classification algorithms based on machine learning have been widely used.With the explosive growth of data scale in the era of big data,the computational resources involving in machine learning algorithms have been increasing persistently.Meanwhile,the development of computer performance is about to encounter a bottleneck.Quantum computing has potential exponential advantages over classical computing and is a significant direction promising to break the performance bottleneck of classical computers.Therefore,combining machine learning with quantum computing is an effective way to make machine learning adapt to the era of big data.In this dissertation,by combining the generative adversarial network or the convolutional neural network in machine learning with quantum computing,the quantum generative adversarial network and the quantum convolutional neural network for image processing were constructed.The specific research contents of this dissertation are as follows.Based on the generative adversarial learning strategy and the remapping method,a quantum generative adversarial network for image generation via learning discrete distribution was designed.The quantum generative adversarial network consists of a variational quantum generator and a classical discriminator.According to the characteristics of classical data generation tasks,the variational quantum circuit design on the quantum generator was simplified to reduce a large number of parameters with a negligible degradation in the performance of the generator model.A discrete probability distribution was converted by the grayscale values of an image,then the probability distribution was approached by the probability amplitude of the quantum state generated by the quantum generator via adversarial learning,and the generated image was obtained from the generated probability distribution.The original complex multimodal distribution of an image can be converted into a simple unimodal distribution by the remapping method.Experiments on the MNIST handwritten digit images dataset and the Fashion-MNIST clothing images dataset indicate that the quality of images generated by the quantum generative adversarial network can be enhanced greatly by the remapping method.Inspired by the convolution and pooling operations of the convolution neural network,a quantum convolution neural network with a variational quantum circuit was designed.Based on the data dimensionality reduction technology,the high-dimensional data of images were mapped into an appropriate low-dimensional space,and the data in the low-dimensional space were encoded into the variational quantum circuit.By taking advantages of angle encoding and amplitude encoding,a hybrid amplitude encoding method with tree structure was presented,which can adjust the width and the depth of the variational quantum circuit required for data encoding flexibly.It is shown that the novel encoding method has a stable encoding performance.By collecting images generated by the quantum generative adversarial network,two artificial datasets for training the quantum convolution neural network were built,and the feasibility of the quantum generative adversarial network for image dataset supplement was verified.The experimental results on the MNIST and the Fashion-MNIST datasets demonstrated the quantum convolution neural networks have faster convergence speed and higher image classification accuracy than the classical convolution neural networks.