Research On Online Super-resolution Reconstruction Of Video Image

Video image super-resolution reconstruction algorithm is an important method to improve the resolution of image sensors.With the increase of hardware cost and demand environment,traditional super-resolution reconstruction algorithms have the disadvantages of low reconstruction quality and limited scope of application.Unlike single-frame image reconstruction,video images need to use the spatio-temporal information of multiple frames,so video image reconstruction is relatively more difficult.Aiming at the problems of occlusion and registration distortion during large-scale motion in video images,to further enhance the reconstruction quality and details of the image,a video image super-resolution reconstruction algorithm based on cyclic image registration was proposed,using non-local convolution Extract image features,integrate image denoising and implicit motion estimation to improve registration accuracy;introduce temporal correlation and attention mechanism to fuse registration features,combine pyramid residual network and parallel convolution to reconstruct high-resolution video images;The combined loss function of pixel loss and texture loss trains the network.Due to the poor visual perception quality and temporal consistency of the current reconstruction algorithms,a super-resolution reconstruction algorithm for video images based on generative adversarial networks is proposed.The generation network uses a lightweight image registration network combined with the global context The information is used to register adjacent images to make up for the missing pixel area,and the global information is extracted using dilated convolution.The improved image discriminative network and the loss of time consistency enhance the visual coherence of the generated image.The full-text information loss is introduced to calculate the similarity of the image during registration distortion Sexual constraint network training.The reconstructed image quality of the proposed algorithm is trained and tested on the Pytorch platform by synthesizing public data sets.The experimental results show that the method based on cyclic image registration can obtain a higher PSNR index,the reconstructed image details are clear,but the visual perception quality is low;The method based on generating adversarial networks can generate realistic texture details,has a high LPIPS human eye perception quality index and good temporal consistency,and the network parameters are fewer and faster than the former.This paper proposes a method that can quickly reconstruct high-resolution images and restore high-frequency information lost in real-world degraded images.