Single Image Super-resolution Method Based On Internal Instance Database And Singular Value Decomposition Optimization

Image resolution refers to the number of pixels in an image,which represents the storage capacity of image information.In daily life,due to the limitations of acquisition environment and equipment or the requirements of data transmission,a large number of valuable images exist in the form of low resolution with low information content.However,single image super resolution algorithms are to use digital image processing technologies,from low-resolution images themselves,to reconstruct high-resolution images containing more image information and details,and finally meet the needs of production and life.Because of its wide application in medical imaging,remote sensing imaging,monitoring and detection,the research of single image super-resolution has always been one of the hotspots in the field of image processing.At present,the traditional interpolation methods are widely used because they are simple to operate and easy to understand.But because of the lack of high-frequency information,they also make the image too smooth.Among the methods based on reconstruction,the iterative back-projection method can effectively compensate for the high frequency information of the image.But at the same time,it introduces some image error without processing.With the rise and development of computer technology in recent years,the sample-based learning methods have become the mainstream of single image super-resolution algorithms.However,for the learning methods based on external image library,the results depend too much on the content correlation between external instance database and target magnified image.In addition,in the learning methods based on internal image library,there are some research problems,such as how to build a portable and efficient internal instance database,and what kind of learning and training model can be used to reconstruct higher quality and high-resolution images.Solving these problems has always been challenging and valuable.Based on the above analysis,this article proposes a new single image super-resolution algorithm based on internal instance database and singular value decomposition optimization.This algorithm mainly includes three parts:learning and mapping based on internal instance database,bicubic surface interpolation based on local features and non-local structure optimization based on singular value decomposition.Firstly,according to the research on self-similarity of the image,this paper uses bicubic subsampling method and our bicubic surface interpolation amplification method based on local features to build an internal instance database from a low-resolution input image itself.Combining the linear regression model in machine learning,it is to learn and map the missing high-frequency image information in the initial interpolated magnified image.The bicubic surface interpolation amplification method based on local features is to use the biquadratic polynomial to fit multiple local surfaces.And the edge information in the image is utilized when solving the polynomial coefficients.Then the bicubic sampling surface of the whole image is obtained by fusing the overlapping local surfaces.Experiments show that the performance of this method is better than that of the traditional bicubic interpolation method.At the end of the algorithm,in order to reduce the image error caused by interpolation and learning,the operation based on singular value decomposition and soft threshold shrinkage is added on the traditional iterative back-projection model,which further improves the quality of reconstructed high-resolution images.In order to verify the rationality and effectiveness of each step in the above image super-resolution algorithm,in the part of experimental results,phased experiments are designed for the following aspects:the comparison between the local feature-based bicubic surface interpolation method and the traditional bicubic interpolation method,the comparison between the singular value decomposition based non-local structure optimization method and the traditional iterative back-projection method,as well as the analysis of the advantages of the internal instance database.In addition,this paper also analyzes and evaluates the overall performance of the proposed method by conducting tests and comparative tests on Set5,Set14 and Urban 100 datasets under different magnification ratios.By comparing with the different types of image super-resolution methods,it is found that the proposed method achieves good results both in visual effect and quantitative indicators.