| The purpose of the image super-resolution is to reconstruct and restore the highresolution representations from the given low-resolution images by using numerical calculation.In many fields,such as medicine,remote sensing and video monitor,the application of image super-resolution has been very common,and how to restore real high-quality super-resolution images has always been a challenging problem.In recent years,due to excellent learning ability and nonlinear representation ability,deep network has been widely used in image super-resolution,which makes the system performance improve continuously.Deep back-projection technology can effectively obtain and enhance the high-frequency features,predict more high-frequency details and improve the reconstructed image quality by alternately extracting and fusing the residual information.Therefore,aiming at the limitation of existing methods that can not fully extract residual features of different scales,which restricts the better improvement of super-resolution performance,this paper studies an image super-resolution method based on dual-path back-projection.It is proposed to enhance the high-frequency features of images at different scales by introducing parallel back-projection technology,so as to improve the quality of reconstructed images.At the same time,aiming at the problem of high magnification super-resolution of image,a progressive learning image super-resolution based on dual-path back-projection is further studied to effectively ease the training difficulty and decrease the computational complexity of high magnification super-resolution network.The work of this paper is as follows:1)Research on an image super-resolution reconstruction network based on dual-path back-projection.In the proposed network,a multi-level dual-path parallel back-projection feature enhancement module is introduced into the unit which extracts the feature,and the residual feature information in different frequency bands is obtained by using up and down sampling in the opposite order on two paths.Using the fusion of multi-level residual feature information,the high frequency features of the image are better enhanced.In addition,the network uses both multi-scale feature extraction and channel attention mechanism to improve feature expression and learning ability.A large number of test results show that the proposed method has achieved good results in improving the performance of super-resolution reconstruction network and reducing the model complexity.2)Research on a progressive learning super-resolution method based on dual-path back-projection.On the basic of work 1,by introducing the idea of progressive learning,the strategy that step by step executes 2× up-sampling operation is adopted,which realizes the effective improvement on the performance of high magnification super-resolution.The proposed network combines the correction feedback mechanism and multi-level jump connection to enhance the feature fusion of multi-scale information.The progressive structure not only significantly reduces the difficulty of network training,but also improves the stability of training.In addition,the method introduces residual dense block and multi-scale discriminant aggregation unit to improve the reconstruction performance of the network.The test results based on multiple public data sets show that the proposed method has effectively improved the reconstruction quality,lower model complexity,and better practicality than the existing similar methods.In general,this paper mainly studies an image super-resolution network design based on back-projection,especially for two problems: how to obtain high frequency feature information of different scales and how to reconstruct the high-quality images at high magnification.Based on these,we respectively studied the image super-resolution network based on dual-path back-projection and the dual-path back-projection based on progressive learning.The proposed scheme shows good results in improving the performance of super-resolution reconstruction network and reducing the system complexity. |
