Research On Image Super-resolution Algorithm Based On Multi-scale Residuals

Image is an important method to record information and also an important carrier for dissemination of the information.There are various reasons for the low image resolution.Image super-resolution reconstruction technology is very effective for the restoration of image pixels.It is of great research significance in the field of small target detection,satellite and remote sensing images,mobile network communications,and video surveillance.The current image super-resolution reconstruction technolog developed rapidly and have achieved good results.However there are still problems such as the too large amount of parameters difficult network training insufficient extraction of low-resolution image feature information,insufficient utilization of information flow in the network,and limiting the learning ability of the network.This thesis improves the existing network model from three aspects: parameter quantity,feature extraction,and upsampling.(1)Aiming at the situation that the network has insufficient extraction of the original input image features and the amount of parameters is too large,an image super-resolution algorithm based on the Multi-scale Attention Residual Network(MCRN)is proposed,which mainly uses three convolutions of different scales to extract image information at multiple levels,replaces the large convolution with a small convolutional group,and controls the number of parameters in the network while increasing the receptive field and nonlinear expression ability of the network.The channel attention mechanism is used to allocate different weights to the network channels,focusing on reconstructing high-frequency effective information.Experiments show that MCRN not only effectively extracts image features,but also reduces the number of parameters.(2)For the channel parameters in the network are not fully shared,the ordinary convolution is not enough for the information extraction ability of the pixel block skeleton part,and the upsampling is not enough for the different magnification factor requirements.This thesis proposes an image super-resolution algorithm based on the Multiscale Asymmetric Residual Adaptive Network(MSRAN),which shares parameters of the information streams extracted between different scales,enhances the extraction of skeleton information with asymmetric convolution,and adaptively reconstructs the upsampled part to enhance the versatility of different magnifications of images.Finally,the proposed network modules were tested on a dataset such as Set5.The experimental results show that for reconstructing different scene images,the algorithm models proposed in this thesis have a certain improvement in both the subjective effect of reconstruction and the objective evaluation standards compared with the algorithm models such as LapSRN and MSRN.