Super-Resolution Reconstruction With Multi-Frame Defocused Images Based On Generative Adversarial Network

As an important research direction of image processing technology,Super Resolution(SR)aims to use hardware or software to reconstruct the corresponding High Resolution(HR)image for the existing Low Resolution(LR)image.Generally speaking,The higher the resolution of the image,the more useful information can be provided,which is of great significance for industrial application and scientific research.In the early stage,the super-resolution reconstruction mainly adopts interpolation or reconstruction algorithms,including nearest neighbor interpolation,bilinear interpolation,iterative back projection,projection onto convex set,etc.In recent years,with the explosive improvement of computing power,deep learning has been developed greatly in the field of computer vision,and the super-resolution reconstruction technology based on deep learning has become the mainstream.This paper will focus on the research status of SR technology,introduce the reconstruction algorithm based on multiple defocus images,and elaborate the innovation and improvement of the classical antagonistic algorithm.The main contents on images super-resolution in this paper is as follows:(1)Based on motion blur and gaussian blur,construct a convolution denoising Auto Encoder(AE)model.The blur kernel generated by random trajectory simulates the motion blur process of image,and the image dataset including motion blur and gaussian blur is used as lossy data to input to the coding part of AE.In the decoding part,through the training of the convolutional neural network,model can predict the original undamaged data as output based on encoded data.The whole process realizes the feature extraction and deblurring process of the image.Finally,the network structure and parameters of the coding part are reserved to lay a foundation for the next step of feature fusion and image super-resolution.(2)Studying SR algorithm based on generating against network Generative Adversarial Nets(GAN)and completing feature fusion with defocused images.By constructing the Spatial Feature Transform(SFT)network layer structure,the encoded and concatenated defocus images can be combined with the generated network framework of GAN through a kind of affine transformation.Then,the whole framework is trained together and fusion more image features.In addition,the relativity discriminator is introduced to improve the original loss function,assist the model to converge and realize the enhancement of the reconstructed image visual effect.(3)Studying the practical application scenarios of super-resolution technology.Two correction preprocessing algorithms are proposed for the problem of different angles of water meter images under complex conditions.Constructing digital dataset and training YOLO detection model to realize accurate number recognition.By analyzing the digital recognition under different resolutions and comparing the accuracy of recognition,it is verified that the super-resolution reconstruction algorithm can be better applied to the related fields of image recognition and has important research value.