Research On Image Super-resolution Reconstruction Based On Deep Learning

Image super-resolution reconstruction(SR)can use one or more low-resolution images to recover high-resolution image,it has become one of the research hotspots in the field of computer vision in recent years.Learning based methods can use a series of additional training samples to learn priori information to reduce the ill-posedness and achieved the best super-resolution result,which mainly consists of sparse coding based method and neighbor embedding based method.However,these methods only seek for sparse coding coefficients and learning the embedding space in low level feature space of the image,which makes it difficult to strictly satisfy the sparseness and manifold assumption,thus leads directly to the degradation of reconstruction result.In this thesis,we mainly do a deep research and analysis on the deep feature representation,neighborhood precision selection and reconstruction objective function for super-resolution task.The specific works are as follows:1.Deep metric learning based SR method.In order to obtain the feature space with higher neighborhood preserving ratio,we combine deep learning and metric learning and take the distance information between high resolution image patches in the training samples as target,thus the new metric space is learned by a deep neural network,in which the neighborhood preserving ratio between image patches is greatly improved.Reconstruction is conducted by sparse multi-manifold embedding in the metric space.We assume that image patches lie in many similar manifolds,by solving a sparse optimization problem,appropriate number of neighbors in the same manifold are obtained adaptively.Due to the introduction of metric learning,the feature space is more consistent with the popular structure of the original space.Experiments show that the proposed method has about 1.5d B increase in PSNR compared with conventional neighbor embedding method.2.Sparse auto-encoder and extreme learning machine based SR method.Considering the high training complexity of the deep learning model,we first use the sparse auto-encoder to extract the high level sparse features of low-resolution image patches,and the fast universal approximation ability of extreme learning machine is used to train the network,then the extracted nonlinear feature is projected to the pixel space of high-resolution image patches to restore high-resolution images.As the advance sparse features can better reflect the essential properties of image patches,the recovered image has more clear edges and rich high-frequency details.Experiments show that the proposed method has achieved excellent results in visual effect and numerical index,and recovered image has clearer texture and richer details.3.Deep residual network based SR method.The objective function of conventional SR methods is mainly focused on minimizing the mean square error.Reconstructed image obtained by solving these optimization problems have high peak signal-to-noise ratio,but lack of satisfactory high frequency information.Considering the excellent learning performance of deep residual network,we design a new deep residual network model for image super resolution reconstruction.This method can effectively utilizes the information in the receptive field,and reducing the difficulty of training.By introducing the loss function generated by the pre-trained VGG19 network,the reconstruction results are visually more real and pleasant.Experiments show that the proposed method is more visually pleasant,and the recovered image is clearer and the details are more abundant.