Research And Implementation Of Super-Resolution Reconstruction Of Massive Images

With the rapid development of the Internet industry,social networking pictures have grown tremendously and people have increased their demand for image clarity and speed.Image superresolution technology has become a research hotspot in the field of artificial intelligence.It has been widely used in medical image processing,facial recognition,surveillance and ultra-high definition television and other fields.As an important research method for image processing at present,deep learning has obvious advantages over traditional methods.There are two problems in the existing image super-resolution algorithms based on deep learning: On the one hand,the traditional convolutional neural network algorithm can not extract the features of the image well during training of the network model.Shallow convolutional neural networks cannot extract better edge and texture information.Although deep stacking convolution can better fit the features of the image and learn higher level texture information,it may also cause gradient dispersion and feature loss due to too many layers;On the other hand,the losses used in the image reconstruction are single-tasking losses,which can only play a certain degree of guidance in the image reconstruction process and it cannot be better to rebuild the image.(1)Aiming at the problem of feature extraction,we mainly study how deep learning network architecture can extract high-frequency texture information from images.This paper proposes an image super-resolution method based on the residual neural network and an experimental comparison and analysis of networks with and without residual and layered structure is performed.(2)Aiming at the problem of image reconstruction,we mainly study how to reconstruct super-resolution images faster and better in deep learning.This paper uses deconvolution instead of interpolation upsampling and tests on the test set of the large-scale image detection contest to prove the effectiveness of the method.(3)Aiming at the problem of network convergence speed,an image super-resolution method based on multi-task loss is proposed.The multi-task loss function in this paper is applied to the two-phase training in this paper.Multi-tasking loss can greatly speed up the training of the network.For the same network structure,using the loss function at the 2× and 4× images respectively to adjust the parameters of the network at the same time than using only a loss function to adjust the parameters at the 4× image output of the network,the convergence speed of the network is enhanced.The learning efficiency of the Internet has been greatly improved.(4)Based on the improved algorithm,this paper designs and implements an image superresolution system based on deep learning.By adopting the programming mode of Python,this article carries on the detailed analysis and the design to each module in the system.The system is mainly divided into data acquisition,data upload and image reconstruction modules.The image super-resolution module uses the MSRCNN-fast,DTLCNN and MCRCNN algorithms proposed in this paper.In summary,compared with many classical algorithms of image super-resolution,the algorithm proposed in this paper has certain advantages in super-resolution and it has also achieved some practical value in practical life.