Image Denoising And Generation Algorithms Based On Quantum Deep Learning

The increasing development of information technology has brought people into the era of big data.Among all kinds of data,image is the most commonly used information carrier in people’s lives.With the large increase in the amount of image data,it is urgent to perform high-performance denoising and generation processing on images to deal with the problems of slow processing speed and poor processing effect.Therefore,image denoising and generation algorithms combining quantum computing and deep learning were designed,respectively.The main work of this dissertation is as follows:Based on convolutional autoencoder and residual learning,an image denoising algorithm was proposed.The encoding layer of the convolutional autoencoder adopted the structure of alternating convolutional layers and max-pooling layers,while the decoding layer utilized the structure of alternating convolutional layers and deconvolutional layers.The nonlinear feature extraction of the input image was completed by the encoding layer and the decoding layer,and a good feature representation was obtained.The residual information was output from the convolutional autoencoder through residual learning,and the proposed algorithm was trained by the cross-entropy loss function.The proposed algorithm was simulated on the MNIST dataset and Fashion-MNIST dataset.Compared with the convolutional autoencoder denoising algorithm without residual learning,the denoised image of the proposed algorithm has better results in both visual effects and objective quality evaluation index.By combining quantum information theory with machine learning method,Born machine generative model based on the probability interpretation of the wave function provides a new tool to study the generative model.The Born machine generative model with the general parameterized quantum circuit generally requires the same number of qubits as the sample feature size of the dataset to be processed,while each sample usually contains thousands of features in actual dataset.To solve this problem,a Born machine generative model with the matrix product state quantum circuit was proposed,which requires less qubits than that with a general parameterized quantum circuit,so it can save the use of scarce qubit resources in near-term quantum devices.And the proposed Born machine generative model was trained with the maximal mean discrepancy loss function.The proposed Born machine generative model was simulated on the Bars-and-Stripes dataset and the mixture of Gaussian distribution dataset.It is shown that the Born machine generative model based on the matrix product state quantum circuit can achieve good generative performance with fewer qubits.