Research On Image Super-Resolution Algorithm Based On Convolutional Neural Network

Image super-resolution reconstruction is to recover high-resolution images from one or more low-resolution images,which is an important task in image processing.This technology has important applications in many fields,such as medical imaging,remote sensing imaging,public safety and old video restoration,etc.In recent years,with the development of deep learning and it has successfully applied in the image super-resolution technology,a large number of image super-resolution algorithms based on deep learning have emerged.However,the number of model parameters of many convolutional neural networks is large and the complexity is high,which leads to difficulties in training and requires a large amount of computational resources and memory consumption.On the other hand,the input low-resolution image information and shallow features of the deep network can not be fully utilized,and the high frequency information of texture details is difficult to recover.To solve the above problems,this dissertation proposes two deep learning-based algorithms to study and improve the super-resolution reconstruction.The main work is as follows:Firstly,a lightweight multi-scale residual network is proposed.The core of the network is designed as a lightweight multi-scale residual module.According to the idea of bottleneck layer,the number and complexity of the whole network are reduced by introducing 1×1 convolution.Secondly,the multi-scale structure in the module contains convolution kernels of different sizes,which is helpful to through different receptive fields to obtain the diversity of image features.In the reconstruction part,multi-scale design is introduced to improve the reconstruction ability.In order to further reduce the parameters of the model,we analyze the influence of different convolution methods on the number and precision of parameters.After that,deep separable convolution is to replace traditional convolution.Secondly,a super-resolution network based on dual-attention mechanism is proposed in this dissertation.The channel attention and spatial attention can learn the different types of features from two different perspectives of channel and space,those two attention methods could better extract and utilize the high-frequency features that are helpful to texture detail recovery more effectively.In addition,a variety of feature fusion methods are adopted to fuse the features of different scales in the proposed basic residual module,so as to further enhance the information flow and feature utilization of the network,which can obtain more effective feature information.Finally,the two types of networks proposed in this dissertation all use local feature fusion modules and global residuals to alleviate the difficulty of network training and accelerate network convergence.In order to verify the effectiveness of the algorithm,tests and evaluations carried out on multiple data sets.The results show that the algorithm in this dissertation has achieved good improvements in objective performance indicators and subjective visual effects.