Multi-Feature Parallel For Image Super-Resolution

Deep convolutional neural networks are widely used in the field of single-image super-resolution,which significantly improves the performance of the reconstruction method.A lot of methods has made notable achievements by growing the network depth.However,most methods use increasing the number of convolutional layers or the number of convolutional layers to improve the performance of reconstructed images.When the number or number of convolutional layers increases,the parameter amount of the model and the memory capacity occupied also increase.In the real world,due to the limitations of hardware devices,the practical application performance of these methods is greatly reduced.This paper investigates and analyzes the current single-image super-resolution method with superior performance,and proposes a potential solution to improve the performance of super-resolution image reconstruction under resource-constrained conditions.First,this paper proposes a Mutil-Feature Parallel Network.The method can obtain the rich hierarchical feature information of the input image,and significantly improve the reconstruction performance by rebuilding different aspects of the super-resolution image.Secondly,this paper improves the Mutil-Feature Parallel Network,and proposes a lightweight Mutil-Feature Parallel Fusion Network for single image super-resolution,which achieves a good balance in reconstruction performance,number of parameters and computation complexity.The main contributions of this article are as follows:(1)Propose a parallel residual block.The feature parallel residual block adopts the strategy of first expansion and then compression.The first and second convolutional layers first expand the channel of image feature information to extract rich image feature information.The third and fourth convolutional layers compress features.Information channel,screening image feature information that is conducive to image reconstruction.The second convolutional layer and the third convolutional layer use 1×1 convolutional layers.Compared with convolutional layers of other sizes.as the number of channels increases,the amount of parameters will not be significantly increased.(2)Propose a correlation weight module.The correlation weight module can calculate the weight of each feature information channel according to the image feature information,and weight each feature information channel separately,so that the feature information that is conducive to image reconstruction can be transmitted.In addition,the adaptive weight module has no parameters.(3)Propose a shallow feature mapping module.The shallow feature mapping module extracts rich shallow feature information through different branches and different scale convolution layers.The rich shallow feature information helps the reconstruction module to reconstruct high-quality images.(4)Propose a multiple reconstruction module.The three-way reconstruction module can reconstruct different aspects of the image with the shallow feature information and the deep feature information,and finally add all the reconstruction results to generate the final reconstructed image.