Research On Image Super-resolution Reconstruction Algorithm Based On Multi-scale Convolutional Neural Network

With the advent of the information society,images have become one of the main ways for humans to obtain and transmit information.Compared with low-resolution images,high-resolution images contain more image feature information and better visual effects.In modern society,the requirements for image resolution in image applications in various fields are increasing.During the process of image acquisition,the environmental factors of the image acquisition,the quality of the hardware equipment for the acquisition of the image,the storage method of the acquired image,etc.,will all lead to the phenomenon that the resolution of the acquired image is reduced.Therefore,the image super-resolution reconstruction technology of obtaining high-resolution images by processing low-resolution images has attracted extensive attention of many scholars.Using convolutional neural networks to reconstruct images with super-resolution is the mainstream method of image super-resolution reconstruction algorithms at present.However,the reconstruction effect of images is more dependent on the model and parameter settings of convolutional neural networks.Aiming at the above problems,this paper proposes two kinds of image super-resolution reconstruction algorithms,which are model optimization and model optimization combined with parameter optimization.The main contents of the image super-resolution reconstruction algorithm based on multi-scale and residual network are as follows:(1)A multi-scale densely connected image feature information extraction network is proposed.The use of multi-scale convolution kernels in the process of image feature information extraction enriches the feature information extraction method of the network model.The multi-scale convolution kernel is densely built,and the diversified image feature information extracted by different scale convolution layers is integrated,which effectively improves the image extraction ability of the overall network,enhances the information circulation in the network,and realize the reuse of channel feature dimension.(2)Residual network can be used to solve the problem of image feature information loss and gradient disappearance in deep convolutional neural networks.The establishment of multiple residual networks can supplement image feature information in multiple aspects and levels,and improve the image reconstruction effect of the overall network model,it effectively suppresses the gradient vanishing problem of weight update in deep convolutional neural network backpropagation.The main contents of the image super-resolution reconstruction algorithm based on multi-scale and attention mechanism are as follows:(1)An improved hybrid convolutional network is proposed.The improved hybrid convolutional network is composed of a combination of traditional convolution and dilated convolution,which avoids the problem of loss of image detail information through a reasonable design of the dilation rate of the dilated convolution.Taking advantage of dilated convolution,the extraction of image contour information is effectively enhanced without losing image details.(2)A multi-scale convolutional and dual-attention fusion network is proposed.Small-scale dense connections and skip connections are designed using multi-scale convolution kernels,and the channel attention mechanism and position attention mechanism are used to optimize the dimension and position parameters of the extracted image feature information.Enhance the feature representation of valid image information,and suppress invalid feature information and interference information.(3)A network module of multi-scale hybrid convolution and attention mechanism is proposed.This module is a reasonable mix of improved hybrid convolutional networks,multi-scale convolutional and dual-attention fusion networks,and conventional networks.The multi-scale convolution and double attention mechanism network is set as the main extraction unit of image feature information,the improved hybrid convolution network and residual network are used to efficiently supplement the image feature information,and part of the activation function is improved to ensure the module Perfect convergence of weights in backpropagation.The proposed module is used to build the overall network model of module stacking.The overall network model of stacking can increase the number of modules one by one,and find the most effective number of modules through testing to achieve the best reconstruction effect.In order to verify the effectiveness of the two algorithms proposed in this paper,this paper compares the proposed algorithm with the selected classical algorithm in experiments.Experiments show that the image reconstructed by the algorithm proposed in this paper is clearer,the detail texture is more perfect,and good results are obtained in both objective evaluation and subjective evaluation.