Image Reconstruction Based On Object Modeling

High-resolution and high-quality images have important applications in many fields such as defense security,game entertainment and medical diagnosis.Due to unavoidable factors such as the hardware limitation of sensor,thermal noise in the circuit system and light con-ditions,the noisy,low-resolution and low-quality images are obtained.High-quality images are reconstructed from the low-quality images via image reconstruction methods including image demosiacking,denoising and super-resolution,whose key is the accurate modeling of the prior knowledge of image objects and noise.In order to obtain the high-quality im-ages,this dissertation starts from two aspects including joint demosaicking and denoising and reducing image noise.Joint color image demosaicking and denoising,mixed noise re-duction,real color image denoising and video foreground estimation have been studied in depth and we had achieved impressing results.The main work and main innovations of this dissertation are as follows:(1)Joint color image demosaicking and denoising(JDD)method via residue learning is proposed.Most of conventional color image demosaicking methods often fail around the ar-eas of strong textures and produce disturbing visual artifacts such as aliasing and zippering.The existing deep learning based methods are capable of obtaining images of better qualities though at the price of high computational cost.To solve the above problem,we propose joint demosaicking and denoising method via residue learning.The design of lightweight dense-ly connected network architecture is inspired by a rigorous analysis of JDD using sparsity models and the success of deep residue learning and aggregated residual transformations.And the network is trained in an end-to-end manner to learn the mapping from the noisy low-resolution space(CFA image)to the clean high-resolution space(color image).Experi-mental results conducted for both demosaicking-only and JDD tasks show that the proposed method performs much better than competing methods(i.e.,higher visual quality,smaller training set and lower computational cost).In addition,the proposed lightweight network is adopted to attack image reconstruction problem such as Gaussian white noise suppression and image super-resolution and achieves the impressing performance.(2)An effective mixture noise removal method based on Laplacian scale mixture(LSM)modeling is proposed.Recovering the high-quality images form images corrupted by addi-tive white Gaussian noise(AWGN)and impulse noise(IN)is a challenging problem due to its difficulties in accurate modeling of the distributions of the mixture noise and the different property of AGWN and IN.Many efforts have been made to first detect the locations of the impulse noise and then recover the clean image with image inpainting techniques from an incomplete image corrupted by AWGN.However,it is quite challenging to accurately detect the locations of the impulse noise when the mixture noise is strong.To solve the above problem,we propose an effective mixture noise removal method based on Laplacian scale mixture modeling and nonlocal low-rank regularization.The outliers caused by IN are modeled with LSM distributions,and both the hidden scale parameters and the outliers are jointly estimated from the observed noisy image through maximum a posteriori method.To exploit the nonlocal self-similarity and low-rank nature of natural image,a nonlocal low-rank regularization is adopted to regularize the denoising process.Experimental results on synthetic noisy images show that the proposed method outperforms most of existing mixture noise removal methods.(3)Real color image denoising based on multi-channel nonlocal low rank model is proposed.Most of the existing denoising algorithms are developed for grayscale images.For color image denoising,the straightforward solution is to apply the grayscale image denoising algorithm to each channel of color image.However,such a straightforward solution do not exploit the spectral correlation among RGB channels and the denoising performance is unsatisfactory.To solve the above problem,we propose real color image denoising method based on multi-channel nonlocal low rank model.Considering of the difference of the noise statistics in R,G and B channels of real noisy images,we introduce a weight matrix to balance the data fidelity of the three channels and hence the spectral correlation among RGB channels has been well exploited.Experimental results on real color image dataset show that the proposed method achieves the satisfactory denoising performance.(4)Robust foreground estimation via structure Gausssian scale mixture modeling(GSM)is proposed.Recovering the background and foreground parts from video frames has impor-tant applications in video surveillance.Under the assumption that the background parts are stationary and the foreground are sparse,most of existing methods are based on the frame-work of robust principal component analysis(RPCA),i.e.,modeling the background and foreground parts as a low-rank and sparse matrices,respectively.However,in realistic com-plex scenarios,the conventional l1 norm sparse regularizer often fails to well characterize the varying sparsity of the foreground components.How to select the sparsity regularizer parameters adaptively according to the local statistics is critical to the success of the RPCA framework for foreground reconstruction task.To solve the above problem,we propose to model the sparse component with a GSM model.Compared with the conventional l1 norm,the GSM based sparse model has the advantages of jointly estimating the variances of the sparse coefficients(and hence the regularization parameters)and the unknown sparse co-efficients,leading to significant estimation accuracy improvements.Moreover,considering that the foreground parts are highly structured,a structured extension of the GSM model is further developed.Specifically,the input frame is divided into many homogeneous re-gions using superpixel segmentation.By characterizing the set of sparse coefficients in each homogeneous region with the same GSM prior,the local dependencies among the sparse coefficients can be effectively exploited,leading to further improvements for foreground re-construction.Experimental results on several challenging scenarios show that the proposed method performs much better than most of existing foreground reconstruction methods in terms of both performance and speed.