Image Super-resolution Reconstruction Based On Convolutional Neural Network

Image super resolution enhancement technology has been widely used in medical imaging,video surveillance,aviation,multimedia and daily life.With the development of deep learning network and the continuous improvement of computer computing capability,image super resolution enhancement algorithm has been continuously improved,from the early SRCNN with only three convolutional layers to the current image super resolution enhancement algorithm based on generative confrontation network.Nowadays,image super-resolution enhancement algorithms are mostly based on generative countermeasures network [1],and image super-resolution reconstruction and its application are gradually becoming mature.On the basis of the generative antagonistic network-related algorithm,researchers are still making relevant optimization,such as sampling method,feature extraction,and the change of the number of convolutional layers,so as to make the training effect of image super-resolution enhancement better and better.In this paper,based on the image resolution enhancement algorithm based on generative antagonism network,and combined with the application scenarios such as personalized customization of ceramic products,further improvement and optimization will be carried out in order to obtain better application effect.This paper focuses on the improvement and optimization of three aspects: one is to improve the image sampling operation;The lower sampling method of gaussian pyramid was used to obtain low-resolution image LR.Compared with gaussian filtering,the training effect and training time were optimized and improved to a certain extent,and the important information beneficial to training was optimized in the retained image.Second,in the algorithm process,the normalization operation is removed;Over the network to generate the type and the implementation process of image super-resolution reconstruction,based on the algorithm of normalized operation by the line test,found the normalized operation in the process of practical training consumes a lot of computing resources,the training time was prolonged,therefore chose to remove the normalized operation,improve the efficiency of the algorithm,shorten the training time.The third is the research and application of the neural network based on the super-resolution of lightweight images.The light image super resolution enhancement algorithm is improved,and the training time and reconstruction effect are balanced and optimized.In addition,based on the team application scenario and tensorflow lite,this paper carried out the transplantation and optimization of the lightweight image super-resolution enhancement algorithm in the mobile terminal,and applied it to the personalized ceramic product customization platform of the author's team.