Design Of Real Time Video Super-resolution Reconstruction System Based On Convolutional Neural Network

Super resolution reconstruction technology is to recover high-resolution image or video with clear texture from low resolution image or video by software algorithm.In recent years,it is a research hotspot in the field of image processing,but there are still some problems that are difficult to solve.For example,most of the networks have the problem of too much computation and too large network parameters to be reconstructed in real time,which makes the super-resolution reconstruction technology limited in the field of video surveillance.Most super-resolution reconstruction algorithms only consider the intra frame feature information,and ignore the inter frame correlation of video sequences.The existing shallow video super-resolution reconstruction network has some problems,such as the weak modeling ability of time-domain characteristics,and the poor visual effect after super-resolution reconstruction.Although the current super-resolution reconstruction algorithm based on deep learning has achieved good results in the field of image quality enhancement,it is still difficult to content the growing demand for image visual effect.This paper makes an in-depth study of the above problems,and puts forward targeted solutions.The main research work includes:1.Based on the residual network,this paper proposes a convolution neural network with multi-scale and multi abstraction to realize feature extraction,and fuses the extracted local features with global features,so as to make full use of the feature information of different scales and different abstractions for super-resolution reconstruction,and proposes a lightweight and high-performance super-resolution reconstruction algorithm.With1920×1080 as high-resolution image,the processor is 2080 TI,the speed of this algorithm is 18 FPS for 3 times super-resolution reconstruction of low resolution image(640×480),and 22 FPS for(480×270)4 times super-resolution reconstruction.At the same time,the new algorithm is compared with the classic lightweight super score algorithm in the standard test set.The PSNR and SSIM are improved,and the visual effect is also improved.2.In view of the problem that most of the existing simple super-resolution reconstruction algorithms are difficult to meet the needs of human eyes for visual effects,this paper combines the image super-resolution network with the image color correction network,and carries out color correction on the 3D look-up table obtained by deep learning training of the super-resolution image data,which makes the super-resolution image look better without affecting the speed(? 10 ms per frame).The visual effect is further enhanced to improve the image resolution and solve the problem that the visual effect is not significantly improved after super-resolution reconstruction.3.Aiming at the problem that image super-resolution can't make use of the inter frame characteristics and the existing video super-resolution reconstruction algorithm is difficult to model in time domain,this paper proposes a new cyclic residual convolution neural network based on cyclic convolution neural network,which uses lightweight residual structure to improve the traditional residual structure,and realizes fast video super-resolution reconstruction It can reduce the network parameters and the amount of computation,and speed up the reconstruction speed.On NVIDIA-RTX2080 TI,the real-time 4 times super-resolution reconstruction of low resolution image is realized,and the speed reaches 24 FPS,which improves the practicability of the algorithm.4.In this paper,a four channel video super-resolution reconstruction monitoring system is built on the ubuntu16.04 system and QT platform.With1920×1080 as high resolution image,the new intra frame super-resolution reconstruction network and color correction network are used to achieve super-resolution reconstruction and color enhancement are realized of 20 FPS with good visual effect.