Search On Single Image Super-resolution Reconstruction Based On Self-learning And External Data

Image super-resolution reconstruction refers to the process of recovering high-resolution images from low-resolution images.It is widely used in HDTV,enhancement of medical lesion images and reconstruction of fuzzy images from monitoring cameras.Image super-resolution can be divided into two types: single image super-resolution and multiple image super-resolution.Single image super-resolution can be reconstructed by only one low-resolution image,which is convenient and fast in practical application and has attracted extensive attention in the field of image processing.Single image super-resolution can be also divided into external data training and self-learning of self-data.In the image super-resolution reconstruction based on external data training,the mapping relationship of high and low resolution images is obtained through the learning and training of a large number of external data.However,the image super-resolution based on external data training ignores the mining of low-resolution image's own information,and the reconstructed image has obvious ringing.In the super-resolution reconstruction based on self-learning of self-data,high-resolution images can be obtained by extracting and mapping a large amount of repeated information of self-image.However,the super-resolution reconstruction based on self-learning of self-data only relies on its own information,and the image details are single and the edges are severely serrated,which restricts the effect of image reconstruction.Aiming at the above defects of super-resolution reconstruction based on external data training and self-learning of self-data,we adopted the method of combining external data training with self-learning of self-data to conduct the following two aspects of research:1.Aiming at the defects of traditional super-resolution image reconstruction,that is,using only external data training or using only self-learning of self-data,we propose the method of combining self-learning in super-resolution reconstruction based on external data dictionary learning to solve the reconstruct problem of single image super-resolution.In the super-resolution of traditional dictionary learning,the image to be reconstructed is used as a low-resolution image in the model for training and mapping after interpolation amplification,resulting in inaccurate description of the image to be reconstructed.The self-learning super-resolution method for pyramid image training only relies on its own information for reconstruction and lacks the guidance of external data training.In order to further improve the the quality of image super-resolution,the image features are gaussian weighted in the process of self-learning,and the rotation strategy is used in the external data dictionary training and solution to reduce the reconstruction error.Experiments show that the proposed algorithm has better reconstruction quality than the traditional algorithm.2.Aiming at the super-resolution algorithm based on deep learning,using only external data training,however,the reconstruction image's own information is not mined.We propose a method combining the traditional self-learning super-resolution algorithm of self-data with the super-resolution algorithm of deep convolutional neural network.Using the method of combining the self-learning of self-data and external data dictionary learning,the reconstruction speed is slow,and the defects of the reconstructed image ringing and the reconstructed image sawing are both serious.The experiment shows that the proposed method overcomes the above defects and achieves better image super-resolution reconstruction quality.Through a large number of simulation experiments,it is proved that the proposed algorithm is feasible and effective.Compared with other super-resolution reconstruction algorithms,the proposed algorithm achieves better results,which provides a new idea for the subsequent super-resolution image research.