Learning Sparse And Deep Representations For Image Restoration

With the development of science and technology,image data is growing at a surprising rate nowadays.However,some unavoidable non-subjective factors often occur and lead to a negative impact on image quality,such as occlusion during image acquisition,noises introduced during the transmission process,image damages caused by improper storage.In order to better process and analyze the image data,it is necessary to first recover the low quality images.Therefore,research on image restoration technology is of great significance.Sparse representation of features can exploit the potential patterns of data as much as pos-sible,thus enabling the development of image restoration technology.At present,due to its neuroscience motivation and mathematical supporting theories,sparse coding has become a mature and effective technology in the area of image restoration.In addition,applying sparse constraints to the data for feature selection,thereby retaining discriminant features and removing redundant features,also provides an important key for the development of image restoration.In recent years,deep representation has become the hottest research direction in the field of computer vision.Deep learning relies on data-driven,and the current explosive growth of data size provides a solid guarantee for the development of deep learning.Deep neural networks can learn potentially complex implicit mappings from large amounts of data,without the need to pre-design a priori knowledge.And the models can process tasks in real time.These advantages make it possible to efficiently handle image restoration tasks.This paper focuses on the data processing methods based on sparse and deep representa-tion.From the aspects of image structure information restoration and image color information restoration,several tasks are studied,including image completion,image denoising and image colorization.The contribution of this paper are concluded as follows:1.We investigate a new method for tensor completion,in which a low-rank tensor approx-imation is used to exploit the global structure of data,and sparse coding is used for elucidating the local patterns of data.Regarding the characterization of low-rank structures,a weighted nuclear norm for tensor is introduced.Meanwhile,an orthogonal dictionary learning process is incorporated into sparse coding for more effective discovery of the local details of data.By simultaneously using the global patterns and local cues,the proposed method can effectively and efficiently recover the lost information of incomplete tensor data.The capability of the proposed method is demonstrated with several experiments on recovering color images,MRI data and video data,and the experimental results have shown the excellent performance of the proposed method in comparison with recent related methods.2.We propose a dynamic dual learning network for blind image denoising.Instead of modeling a sole task prediction network,the proposed DualBDNet investigates the inherent relations between the residual estimation and the non-residual estimation.The proposed network produces task-dependent feature maps,each part of the features are devoted to one specific task?residual/non-residual mapping?.To settle different noise levels with a single network or even the case that the statistics of noise are unknown,we further introduce an embedded subnetwork into the proposed model.One output of the subnetwork is to learn a dynamic compositional attention to highlight those more significant task-dependent feature maps,coincided with the extent of corruption adaptively.The other one is to learn a weight used for our results fusion,so as to ensure an end-to-end manner.Extensive experiments demonstrate the proposed method outperforms state-of-the-arts on either synthetic or real noisy images,without estimating the noise levels as input.3.We propose to convert a color image into a grayscale image that can fully recover its original colors,and more importantly,the encoded information is discriminative and sparse that save storage capacity.Particularly,we design an invertible deep neural network for color encoding and decoding purposes.This network learns to generate a residual image that encodes color information,and it is then combined with a base grayscale image for color recovering.In this way,the non-differentiable compression process?e.g.,JPEG?of the base grayscale image can be integrated into the network in an end-to-end manner.To further reduce the size of the residual image,we present an L₀regularization layer to enhance the sparsity,and thus leading to the negligible storage space.Extensive experiments demonstrate that the proposed method outperforms state-of-the-arts in terms of image quality and tolerability to compression.