Research On Image Super-resolution Based On Self-learning And Convolutional Network

Image super-resolution reconstruction method is a process of restoring high-resolution image from a series of related low-resolution images by means of hardware or software.At present,machine learning and deep learning are two better methods for image reconstruction quality among the methods to achieve super-resolution image reconstruction.Therefore,this paper studies the image reconstruction algorithm based on these two methods.Self-learning image super-resolution algorithm based on neighborhood embedding is a machine learning algorithm.Super-resolution reconstruction network for extracting deep image features is a deep learning based algorithm.The main contents of this paper are as follows:(1)Super-resolution algorithm for self-learning image based on neighborhood embedding.The problem of super-resolution image reconstruction based on machine learning is how to accurately describe the mapping relationship between high/low resolution images.Neighborhood embedding algorithm has provided an effective solution to this problem.But how to quickly find the suitable matching patch in the neighborhood embedding is also a difficulty.To solve this problem,this paper proposes a strategy of random oscillation + horizontal propagation + vertical propagation to achieve super-resolution image reconstruction.Firstly,the testing patch was matched to the best matching patch by random oscillation.Secondly,the best matching patch is propagated by the similarity of images.Finally,in the construction of the image pyramid,it is found that there are structural similarities between different image layers in the vertical direction of the image pyramid,so vertical propagation is added in the horizontal four-way propagation.Experimental results show that compared with some self-learning based methods,the proposed method has better reconstruction results.Compared with the external learning based algorithm and the latest deep learning based algorithm,the proposed algorithm has lower time complexity,lower requirements on the storage space of the device,and can better recover the image details.(1)A super-resolution reconstruction network for extracting deep image features.The image super-resolution method based on deep learning is faced with the problem of how to solve the under-fitting problem caused by insufficient training samples.In order to solve this problem,this paper proposes an enhanced feature extraction network architecture.Firstly,the enhancement unit and residual unit are designed in the network to extract the deep features of the image,which can effectively avoid the loss of image information.Secondly,the input images are used as training samples and test samples,and the internal self-similarity of the images is fully utilized to achieve super-resolution reconstruction of a single image.Experimental results show that the proposed network structure effectively solves the problem of under-fitting caused by insufficient training samples,and obtains high resolution images with good visual effect.In summary,this paper uses the machine learning and deep learning to study a self-learning image super-resolution algorithm based on neighborhood embedding and a super-resolution reconstruction network that extracts deep image features.These two algorithms are used to achieve super-resolution reconstruction of the image.