Research On Image Restoration Methods Incorporating Prior Models And Deep Learning

With the development of imaging technology and the popularization of imaging devices such as smartphones,high-quality digital images play an important role in all aspects of human’s production and life.However,in the process of image acquisition,transmission and storage,due to the influence of natural environment,equipment performance and human factors,the obtained images will have problems such as noise pollution and low luminance.Therefore,as important issues in the field of image restoration,image denoising and low-light image enhancement have attracted extensive attention from researchers.In the recent few decades,thanks to the rapid development of deep learning technologies,deep learning-based image restoration methods are significantly outperform traditional model-based methods in terms of inference speed and restoration performance.Most existing deep learning-based methods treat image restoration as learning the mapping function low-quality degraded images to high-quality clean images,and construct end-toend black-box neural networks.However,these methods ignore the physical degradation models of the image restoration problems,and the proposed neural network structure lacks interpretability,which is not conducive to the improvement of the neural network structure,and also hinders the further improvement of the image restoration performance.This thesis mainly focuses on image denoising and low-light image enhancement tasks,and utilize the optimization model and physical model of image restoration respectively to provide theoretical guidance for the design of end-to-end neural networks.Thus,the thesis could improve the interpretability while maintaining the efficiency of neural networks,and enhance the robustness and performance of image restoration algorithms to real-world degradation methods.The main research works of this paper is summarized as follows:1.The thesis first studies the design of the optimization model-driven neural networks.For mixed noise removal task,most existing methods cannot effectively fit the residual of mixed noise images.The thesis proposes a fidelity and regularization terms jointly learning objective function for mixed noise removal,and utilize deep neural networks to learn the distribution of mixed noise and deep prior information of natural images.Then,the proposed image denoising objective function is solved by the half-quadratic splitting method.Finally,the thesis constructs an end-to-end interpretable neural network according to the iterative optimization steps.Experimental results on mixed gaussian-impulse removal task demonstrate the effectiveness and efficiency of the proposed method.2.To further improve the utilization of neural networks in deep unfolding methods,we propose an image denoising method that employs deep unfolding in the feature space.Based on the optimization model of the image denoising problem,we apply its gradient descent iterative process to the deep feature maps and construct a feature-based denoising module that directly removes noise in the deep feature space,thereby reducing the information loss during transformations between deep feature space and raw image space.Besides,in order to enhance the structure and edges in the denoising results,we also propose a multi-scale regularizer module to explore rich prior information from feature maps of different resolutions.Extensive experiments show that the proposed method achieves favorable results on gaussian denoising,blind gaussian denoising as well as real noise removal tasks.3.For low-light image enhancement,the thesis combines the physical model of image enhancement with deep neural networks.Based on the Retinex representation model of images,the theis proposes an illumination-guided convolutional neural network for low-light image enhancement.Given a low-light image,the model first estimates the illumination information from the input image,which represents the degree that each pixel needs to be enhanced,and then uses the frequency feature transformation module to integrate it with the deep image features.Thus the model could enhance low-light images with the guidance of illumination information.In addition,the thesis also proposes an attentive wavelet transform module by combining the attention mechanisms with wavelet transform.The attentive wavelet transform module utilizes channel and spatial attention mechanisms to adaptively modulate the wavelet coefficients of different frequencies,which is beneficial in noise suppression and edges preserving.Compared to existing low-light image enhancement methods,the proposed method achieves better quantitative and qualitative results on synthetic and real-world low-light image datasets.