Image Super Resolution Reconstruction Based On Deep Learning

As a medium,image is a tool for human beings to understand,grasp and represent the world.It is the most effective media in today's society.However,in the process of image acquisition,because of the interference of external factors,the image will be more or less affected.These lead to the image can not meet people's needs and hinder peopled life and work.To improve this problem,relevant researchers have invested a lot of energy in improving image quality.Image super-resolution reconstruction based on deep learning is the best algorithm in image quality reconstruction at present.However,the current deep learning reconstruction algorithm can not satisfy the real-time and high efficiency.To solve this problem,the combination of image reconstruction and deep learning has been studied in depth.The main work content is as follows:First of all,the current research status of image reconstruction at home and abroad is reviewed.Reconstruction-based reconstruction method and learning-based reconstruction method are studied.According to the comparative study of the two methods,the superiority of the super-resolution algorithm combined with deep learning in the current image reconstruction field is summarized.Next,an image super-resolution reconstruction algorithm based on multi-layer residual convolution network is proposed.The method is based on a multi-layer convolution network,established in conjunction with the characteristics of the residual structure.In the process of reconstruction,while increasing the network depth,in order to avoid the computational loss caused by too many parameters,a method of using 1 × 1convolution kernel to reduce the data dimension after convolution is proposed.This method not only speeds up the calculation,but also reduces the loss of information.A multi-scale parallel convolution structure is added to the image reconstruction layer to reconstruct features of different scales to better restore image detail information.Finally,in this thesis,an improved watermarking image reconstruction model is proposed to solve the problem of low watermarking reading rate caused by external factors in trademark image recognition of watermarking information.Based on the deep learning image reconstruction,the model is improved according to the purpose of reconstructing the watermark image.The input image and the label image respectively use a real watermark trademark image with a lower read rate and an electronic version of the corresponding trademark image.And the step of increasing the pixel value is canceled,because for the acquisition of the trademark watermark information,the pixel value increase of the image can ouly play a negative role.The experimental results show that the proposed reconstruction model based on multi-layer residual convolution network has achieved good performance in both reconstruction and practical applications.