| With the rapid development of multimedia technology and the continuous improvement of people’s living standard,people’s demand for high-resolution images in many fields is increasing.The limitations of hardware equipment and the influence of external environmental factors make it impossible to obtain high-quality images.Low-quality images cannot provide us with enough information,and therefore cannot meet people’s application needs.It is generally expensive to improve the spatial resolution of an image by upgrading hardware equipment.In order to reduce the cost,the use of software technology of super-resolution reconstruction algorithm to improve the spatial resolution of the image has become the main method.Super-resolution reconstruction technology has been widely used in many fields,such as video surveillance security field,satellite remote sensing imaging field,video perceptual compression transmission field and medical image processing field,so it has more important research and application value.With the continuous improvement of computer hardware performance,machine learning has also developed rapidly.The learning-based image super-resolution reconstruction method trains the model through a large number of external data samples,trains complex mapping relationships,and reconstructs high-resolution images through this mapping relationship.This paper mainly studies image super-resolution reconstruction algorithms based on learning,and proposes a single-image fast super-resolution reconstruction method of multi-frequency convolution sparse coding and a method of generating super-resolution images based on adversarial edge learning models.The main work of this paper is as follows:(1)This paper briefly introduces the research background and significance of image super-resolution,lists the application scenarios of image super-resolution reconstruction and summarizes its development process and research status;Then it briefly introduces the image degradation model and image quality evaluation standard.According to the different algorithm principles,the reconstruction algorithms based on interpolation,reconstruction and learning are analyzed and generalized.(2)A single image fast super-resolution algorithm model of multi-frequency convolution sparse coding is proposed.Through multi-frequency decomposition,the image is decomposed into residual components and smooth components in multi-scale.The residual components reflects the texture structure of the image,and the smooth components reflects the general picture of the image.The residual components in different frequencies are used to reconstruct the high-resolution image with richer texture details.At the same time,the filter on each frequency is trained in a separate way,which can greatly reduce the burden of computer memory consumption.The experimental results show that the proposed algorithm has better reconstruction effect than the original algorithm and some classical algorithms,and it has improves the reconstruction efficiency.(3)A method of generating super-resolution images based on the adversarial edge learning model is proposed.Most of the image super-resolution methods based on convolutional neural network adopt end-to-end mode,which often have problems such as blurred edges of reconstructed image texture and missing high-frequency information.On the basis of SRGAN,a deep network model based on adversarial image edge learning is proposed,which makes full use of the image edge information to guide the super-division network to generate more realistic high-resolution images.The experimental results show that the proposed algorithm has better performance in peak signal-to-noise ratio,structural similarity and perception indicators than some classic algorithms. |
