| The image super-resolution is an image processing technique that improves the visual effects of existing low-resolution images.The goal of the image super-resolution is to reconstruct a high-quality image with high quality from one or more low-resolution observations.With the development of machine learning and pattern recognition technology,the example based super-resolution algorithm has received a lot of attention.The super-resolution methods based on mixed sample learning are newly developed and it is one of the acheivement algorithms for super-resolution.This thesis takes the fusion of external samples and self-samples as the starting point.The non-redundant characteristics of mixed samples are exploited.This thesis studies the image super-resolution algorithm based on non-redundant information learning with the sparse representation theory and the deep learning methods respectively.The sparse representation,dictionary learning and convolutional neural network theory are introduced.The content mainly includes three aspects.The first part introduces the image degradation model and then the sparse representation theory is presented.The second part introduces two different types of shallow dictionary learning methods: PCA classification dictionary learning methods and K-SVD dictionary learning methods.And their advantages and disadvantages are summarized.The last part introduces deep neural network theory,which includes the Convolutional Neural Networks(CNN)and the Sparse Convolutional Neural Networks(SRCNN).The image super-resolution process is a serious and ill-posed pathological problem due to the complex defect information and the unknown degradation model in the process of low-resolution image generation.The image super-resolution algorithm based on mixed examples and sparse representation is proposed.First of all,the idea of mixed sample selection is introduced,and the corresponding training method is selected according to the non-redundant characteristics of the mixed sample.In the dictionary training stage,a number of compact non-redundant subdictionaries are trained by K-SVD dictionary training methods.And then the orthogonal matching pursuit(OMP)algorithm is taken to select the sub-dictionaries that are most similar to each input image patch.The sparse representation coefficients of each input image patch are calculated for reconstructing the high-resolution image.Finally,the experimental setup and experimental results are given.Due to the restriction of low-level learning methods and the lack of details in high resolution images reconstructed by shallow learning methods,the image super-resolution algorithm based on mixed examples and deep convolutional neural networks is proposed.The residual learning strategy is used to learn the residuals between high and low resolution samples.At the same time,detail learning is introduced.The mixed example selection scheme is proposed.The addition of the mixed example increases the non-redundancy of the training set.Batch normalization and residual learning are integrated together to speed up the training process and improve super-resolution performance.In addition,extensive experimental results show that the proposed method can not only produce good image super-resolution performance quantitatively and qualitatively,but also make the running time shorter. |
