Research On Image Super-resolution Algorithm Through Ridge Regression

High-resolution images are widely used in clinical medicine,computer vision,space science and other fields for its clear image details.However,because of the limitation of hardware equipment,imaging technology,environmental interference factors and network bandwidth,the resolution of images we obtain is often not ideal and cannot reach the required resolution.It often requires high cost to improve image resolution with hardware,so the technology to improve image resolution with software,namely image super-resolution technology,is born at the right moment.The main research in this paper is the technology of obtaining high-frequency information of images by digital image processing technology based on an input low-resolution image,namely the single image super-resolution reconstruction technology(SISR).At present,SISR methods can be generally divided into three categories: image super-resolution based on interpolation,image super-resolution based on reconstruction,and learning based image super-resolution.The interpolation based methods usually use the values of neighboring known pixels around the current pixel to be interpolated to estimate the value of current pixel,which is simple and fast.The reconstruction based methods reconstruct the high-resolution image according to the image degradation model and some prior knowledge.The learning-based methods learn the mapping relationship between high resolution blocks and low resolution blocks in the off-line training phase.In the on-line reconstruction phase,low-resolution block and its closest mapping relation are used to construct the corresponding high resolution block.Some improvements on the original algorithms based on ridge regression are made in the proposed two image super-resolution methods.In the first proposed method,the traditional ridge regression super-resolution framework is also used.In the off-line training phase,the iterative back-projection is used to enhance the features of the B-cubic interpolated images.And the enhanced image features are used in the subsequent ridge regression training.In image reconstruction phase,none-local means(NLM)with gradient factors added is used to strengthen the image after bi-cubic amplification,and then image features are extracted.When searching the mapping function for each image block,unlike the traditional method,which only selects the closest function,a number of optimal mapping functions with weighting coefficients are selected to obtain the corresponding high frequency(HF)details.In this method,whether the non-local means method with gradient factors added to enhance image features or the selection of several optimal mapping functions,They all keep more valuable image information and mapping information,leading to better quality of high-resolution images.The second proposed method also takes traditional ridge regression method as the basic framework.Since the high frequency details obtained by traditional ridge regression based image super-resolution methods may not be ideal and the non-linear relation is ignored,the form of multi-layer ridge regression is proposed.For the consideration of time consumption and effect,the number of layer is finally set to2.Therefore the training stage is also divided into two layers.The high resolution image obtained from the first training layer is used as the initial low resolution image in the second training layer for the mapping matrix training of the second layer.Similarly,the online reconstruction phase is divided into two layers.For an input low resolution image,a high resolution image is reconstructed by the mapping matrices which were trained in the first training layer.Then,another better high resolution image is reconstructed by the mapping matrices obtained in the second training layer.Finally,back-projection method is used to enhance the reconstructed image.The proposed method uses multi-layer mapping to overcome the defects caused by one-layer linear mapping,leading to better quality of the high frequency details of the reconstructed high-resolution image.