Research On Image Super-resolution Algorithm Based On Depth Neural Network

As a classic problem in the field of image,image resolution has received more and more attention in recent years.The resolution of an image determines the amount of information contained in an image,and directly determines the comfort of the image observed by the human eye.Nowadays,with the development of related technologies in machine learning and computer vision,areas such as image classification,object detection and face recognition require high-quality clear images for subsequent professional operations such as image recognition and segmentation.Images with high-frequency information and sharp edges are useful for extracting useful information.The way to improve image quality by means of algorithms is more flexible and less costly than improving hardware.The arrival of the era of big data and the development of artificial intelligence have made image super-resolution algorithms based on neural networks a research hotspot.The research content of this paper is mainly reflected in the following aspects:(1)A super-resolution algorithm based on multi-scale convolutional neural network is proposed,which avoids the dependence of traditional methods on the prior knowledge of low-resolution images.The algorithm is innovative on the convolution kernel "scale" of the neural network model,which makes the algorithm not only can obtain many features of the feature map in the same layer,but also increase the nonlinear expression ability of the network.In addition,the algorithm uses a deconvolution structure to avoid interpolation preprocessing operations before training.The experimental results show that the reconstruction results of this method have certain advantages.(2)A super-resolution algorithm for improving extremely deep neural networks is proposed.The algorithm focuses on "depth" and the internal residual structure and deep structure make the algorithm obtain a larger receptive field while fast convergence.Compared with the improved VDSR algorithm,the proposed algorithm has a deeper network structure,and finally refers to the sub-pixel convolution structure to output the reconstructed high-resolution image.The reconstructed image exhibited by the algorithm has richer texture details than other algorithms and has also greatly improved in objective indicators.(3)In this paper,a comprehensive analysis and experiment are carried out on the two algorithms of "scale" and "depth",and both the subjective feelings and the objective indicators provide a favorable proof of the superiority of the algorithm.We further design experiments to prove the influence of multi-scale structure and deep-level structure on the image super-resolution reconstruction results,and provide theoretical reference and data basis for the future research of neural network in image super-resolution reconstruction.