Optimization-and-Learning Based Image Coding And Enhancement

Using artificial intelligent techniques for image representation,image compression and image enhancement has become a very significant research topic in the fields of image processing,pattern recognition and computer vision,etc.In this thesis,the compression and enhancement of digital images are taken as the study subjects,and we deeply study several key techniques of compression-oriented image representation,standard-compatible multiple description image coding,depth learning-based multiple description image coding and image enhancement,etc.The research works mainly include several parts as follows:(1)We propose a virtual codec supervised image re-sampling compression method,which is used to resolve the non-differentiability problem of the hard quantization function.This method is also extended to multiple description image coding,and we propose a JPEG standard-compliant multiple description image coding method,in which the generated description images are treated as the opposing image labels for each other,while multiple description distance loss is leveraged to effectively constrain the learning of multiple description generation network.Many experimental results verify the effectiveness of the proposed method.(2)We propose a deep learning-based multiple description image coding framework,which can well avoid complex index assignment problems of multiple description quantizers.Firstly,the proposed framework is built upon auto-encoder networks,which are composed of a multi-scale dilated convolutional encoder network and a multiple description residual convolutional decoder network.Secondly,a pair of scalar quantizers and corresponding importance indicators are obtained by an end-to-end self-supervised learning.Testing on several common standard datasets,experimental results show that the coding performance of the proposed multiple description coding method exceeds that of many existing multiple description methods,especially at low bit-rates.(3)Since down-sampling can greatly reduce the amount of data transmission of 3D videos,but the 3D videos are required to be up-sampled at the decoder,we propose a generative adversarial network conditioned by low-resolution color images and depth images to resolve the problem of simultaneous super-resolution in 3D videos.In addition to the adversarial loss,three auxiliary loss functions are used to train the generative network.The experimental results show that the proposed method can generate high-quality images and can be used to resolve other image processing problems.(4)We propose a local activity-driven anisotropic diffusion model and introduce two novel edge-stop functions to efficiently remove severe artifacts in compressed depth images.Meanwhile,we propose a local activity-driven relative total variation model for image denoising and image smoothing.Since the second model is non-convex,so the regularization term of the proposed model is decomposed into quadratic and non-linear parts for solutions.A large number of experimental results show that the proposed method outperforms the existing methods regarding image quality enhancement and image smoothing,and the proposed method benefits to improve image coding efficiency.