Research On Single Image Super-resolution Based On Multi-scale Progressive Reconstruction Method

Image super-resolution technology typically refers to a computer vision task of recovering a high-resolution image with sharp edges,reasonable details and clear contours from single or multiple low-resolution images.In this paper,we mainly focus on single image super-resolution task,which reconstructs the corresponding high-resolution image from a single low-resolution input image.Since high-resolution images can provide a wealth of detailed and effective information,this technology has been widely applied in many fields such as medical imaging,public security,remote sensing,and high-definition television.In addition,it has also become the basis of many computer vision applications such as image classification,target recognition and tracking,image enhancement.Since a single image contains scarce high-frequency details,how to reconstruct a high-resolution image for large scale factors has attracted widespread attention.Aiming at this problem,we propose an improved image super-resolution network based on multi-scale progressive reconstruction to fully exploit the image features and improve the representation capabilities of the network.The direct image reconstruction structure that is employed in many methods,which performs the up-sampling operation in one-step,substantially increases the learning difficulty for larger scale factors(such as 4 or 8).In addition,the required network depends on the scale factor,that is,the trained network only produces super-resolution image with a scale factor.However,inspired by the idea of Laplacian pyramid structure,the proposed network employs the progressive reconstruction network architecture to extract abundant local and global feature information.The structure consists of cascaded convolutional neural networks that realize the process of gradually reconstructing the specified super-resolution image from small scale factor to large scale factor.This structure decomposes a difficult problem into multiple simpler sub-problems,thereby reducing the learning difficulty and improving the effect of image reconstruction for large scale factors.In addition,this paper also proposes an improved multi-scale feature extraction block to achieve deep exploitation for low-resolution image feature information,which effectively improves feature diversity and representation capabilities of the model.And weight normalization is applied in the multi-scale feature extraction block to improve the stability of model training.Finally,this paper adds a pyramid pooling layer into the up-sampling module to further enhance the image reconstruction performance by aggregating local and global context information,so that the reconstructed image can be more visually consistent with the human perception system.In this paper,we use the DIV2 K dataset as the training set,and choose these five publicly available datasets: Set5,Set14,BSDS100,Urban100 and Manga109,as the testing sets.The extensive evaluations on benchmark datasets verify that the improved network proposed in this paper can not only reconstruct more accurate images for larger scale factors,but also be closer to real images in visual effects.In addition,the training stability of model training is also improved.For the scale factor of 8 times,the results of the improved network on each dataset are 27.39 dB,25.11 dB,24.86 dB,22.84 dB,25.00 dB...