Study On Super-Resolution Image Restoration Method Based On The Dictionary Learning And Sparse Representation

With the rapid development of multimedia technology and networking technology,the image information is more and more important in people's work,study and life.But,image will be influenced by technical factors,weather factors and other factors in all aspects of the acquisition,storage and transmission process,causing some inevitable image quality problem which shows explicitly as image indistinct,distortion,noise and so on,and we call this phenomenon the image degradation problem.However,what we need is the clear and high resolution image.It plays a very important role in the application and life for image restoration.Since making a change on image quality through the hardware method is limited by many factors and the higher and higher cost,people began to consider improving the image quality through the software method.Therefore,super-resolution image restoration technology is employed.Now,the research on super-resolution image restoration technology has reached a new level,and this technology is more and more widely applied in various fields.The learning-based image SR restoration algorithm is the most important for image SR restoration algorithms.The characteristics of the learning-based algorithm for image reconstruction are accurate,a strong retention of image features and the strong robustness of image noise.Based on the framework of learning,the image SR algorithms of dictionary learning and sparse representation are studied to achieve the ultimate goal of improving subjective visual effect of the image in this paper.The main contents are as follows:The image super-resolution reconstruction algorithm based on dictionary learning of image content and sparse representation introduces the concept of clustering according to the difference between the content of the training images,this cluster based algorithm divides a large and comprehensive dictionary into different categories,the method also makes a targeted classification according to the content of the image to be restored,and this makes the algorithm more distinguished and targeted and makes the image have a good adaptability.Compared with the traditional super-resolution restoration algorithm,this method has certain advantages on objective test data of PNSR,SSIM et al and subjective visual effect.Another image super-resolution reconstruction algorithm is based on dual dictionaries learning of image content and sparse representation.On the foundation of classifying the training images according to the content of the images,the high-frequency information of an image is divided into the main high-frequency information and redundant high-frequency information,we train the dictionaries respectively to form the dual dictionaries,then we operate two-level super-resolution image restoration operations on the original image,the reconstructed image can obtain more abundant high-frequency information,so that the reconstructed image has better effect.Experimental results show that this method can obtain more image details,this method can get better effect compared with some other traditional super-resolution image reconstruction algorithms,and the effect is fully reflected in the experimental data and experimental visual renderings.We research the very low resolution face image super-resolution based on DCT.With the extensive application of the surveillance cameras,the very low resolution problem happens in many applications of face recognition,which has brought great inconvenience to users.In order to research this problem,this paper proposes the improved DCT algorithm,and our final experimental results has compared with Cubic B-Spline interpolation algorithm and Baker et al's method.Experimental results show that the improved DCT algorithm can get better performance in the very low resolution face image super-resolution reconstruction.We put forward a novel double levels of face hallucination framework super-resolution by learning-based and discrete cosine transform.The problem is formulated as inferring the DCT coefficients which contains DC coefficient and AC coefficient in frequency domain instead of estimating pixel intensities in spatial domain.And take the AC coefficient of training images form training set which reconstruct the face image.It not only reduces the size training set,but also better and more simple than the traditional learning-based algorithm,reduce the difficulty and improve the efficiency of algorithm.On this basis,the high frequency information are combination of two components that is the main high frequency and residual high frequency.The dictionary which be trained combined with sparse representation to the face image secondary reconstruction.The method in this paper captured more high-frequency information and enhance the image quality further in face image reconstruction.