Research On Image Super-resolution Algorithm Based On Linear Mapping

High-resolution images generally retain a large amount of high-frequency information,which is convenient for subsequent research and processing.However,due to the limitations of low-performance image acquisition equipment and harsh external environments,it is a difficult challenge to obtain high-resolution images in real-world applications such as remote sensing imaging,video surveillance,and medical imaging.Image super-resolution technology has developed into an effective way to reduce the excessive cost of image acquisition equipment and obtain high-resolution images from multiple environments.This paper mainly studies the single-image super-resolution method,and combines effective prior knowledge to constrain the input low-resolution image to obtain a high-resolution image that retains a lot of high-frequency details.Many existing super-resolution methods can be summarized into three types:interpolation-based,reconstruction-based,and learning-based.Learning-based super resolution methods have become the mainstream of the current research.These methods map the images in the low-dimensional space to the images in the high-dimensional space by learning the mapping relationship between the two training sample spaces of low resolution and corresponding high resolution.Under the premise of different training sample extraction methods,learning-based methods can usually be divided into two categories: one is to use external natural image data sets to extract external training samples,and the other is to use the information of the image itself to extract internal training samples.In view of the difference between the above two methods based on external training samples and internal training samples,this paper proposes two image super-resolution reconstruction algorithms based on linear mapping relationships:(1)A new mapping matrix for enhancing the mapping relationship between low and high resolution sample space is proposed.First,the uncorrelated features between dictionary atoms are added to the training dictionary model so that each dictionary atom is independent of other dictionary atoms.Second,find the nearest neighbor sample set associated with each dictionary atom in the low-resolution and high-resolution training sample sets according to the correlation,and then use the exponential function to calculate the distance between the nearest neighbor sample and its center sample as the marker value.Finally,a new mapping matrix is obtained by combining the high and low resolution nearest neighbor sample set and the label value matrix to obtain a new mapping matrix,so as to reconstruct an image containing a lot of high-frequency details.Experiments show that the numerical results of this method on PSNR and SSIM are better than existing methods.(2)Because the image super-resolution algorithm of the external training samples sometimes affects the reconstruction effect due to the structural dissimilarity between the test image and the training image to a certain extent,an image super-resolution algorithm based on its own samples and weighted representation coefficients is proposed.First,an internal training image set is constructed for a single input image rotation and a horizontal mirror transformation operation,and the training samples are randomly selected as anchor points to obtain high and low neighbor image block sets.Secondly,according to the similarity between the high and low neighboring image blocks,respectively calculate the low and high resolution similarity adjustment weight matrix,and use the two weight matrices to weight the representation coefficients successively.Finally,the mapping matrix between the two training sample spaces with high and low resolution is obtained by regression learning using the weighted representation coefficients and the neighbor image block set,and finally the high resolution image is reconstructed.Multiple experimental results show that this method is superior to many existing methods in visual and quantitative indicators.