Research On Single Image Super-Resolution Using Multi-Scale Information Processing Networks

Single image super-resolution is one of the fundamental issues in computer vision,which aims to generate corresponding high-resolution images from low-resolution images.This technology has wide applications in the fields of medical imaging,object recognition,video surveillance,and remote sensing imaging.Convolutional neural network technology has dominated the current research on image super-resolution,but its performance heavily depends on the size of the network,and the increase in the amount of parameters and computation limits the application of the model on mobile devices such as mobile phones.Lightweight convolutional neural networks refer to convolutional neural networks that maintain high accuracy,have small parameter and computational complexity,and have fast reasoning speed.They are widely used in scenarios such as mobile devices and embedded devices.In order to further improve the performance of lightweight image super-resolution,this paper proposes a super-resolution reconstruction algorithm using multi-scale information processing networks,and constructs a new network model using the idea of multi-scale information fusion.In the network,a new channel reorganization aggregation convolution unit is introduced and used as a basic convolution unit in the network.It can enable each layer of the network to have different sizes of receptive fields and capture and fuse multi-scale features without introducing additional parameters.In the nonlinear feature mapping section of the network,a multiple cascaded residual block is proposed,which enables features of different scales to interact multiple times.To capture non-local feature information,a multi scale nonlocal attention block is designed.The feature map is divided into regions of equal size in the spatial dimension,and then spliced and fused in the channel dimension to obtain richer non-local features.In order to analyze the importance of feature information obtained at different depths,a new multidimensional attention block was designed during feature reconstruction.Pixel importance analysis was performed on the output features of each layer in both channel and spatial dimensions to obtain pixel-level attention weights in different dimensions.This article also proposes a lightweight image super-resolution network based on non local region information complementarity,and the network architecture is improved on the model proposed in the previous paragraph,and the most important one is the improvement of non-local attention blocks.After dividing the feature map into regions of equal size in spatial dimensions,for each region,calculate the similarity between other regions and this region,and assign different weights to different regions based on the similarity.Based on this idea,a new region information complementary attention block is designed,which can enable models to use similar non local region feature information for local region feature reconstruction,making full use of image self-similarity.Experiments have shown that the method proposed in this paper achieves a better balance between model size and performance compared to some current representative methods,and the generated images have better objective and subjective effects.