Research And Optimization Of Video Preprocessing Algorithm Based On Deep Learning

With the popularization of smart electronic products,video has become an important means for people to obtain information,and the quality of video images will directly affect the user experience.The video preprocessing algorithm can recover and enhance the captured video images,so it can provide users with a reliable guarantee for obtaining high-quality video pictures.This thesis is a research on intelligent visual aid technology for visually impaired people.Take a snapshot of the video.First,the super-resolution image reconstruction algorithm is used to enlarge the captured image.Then,the image denoising algorithm is used to improve the quality of the enlarged image.Finally,the image warping correction algorithm is used to correct the denoised image.Provide high-quality and reliable guarantee for the subsequent text recognition system.At the same time,deep learning has achieved great success in the field of computer vision.Therefore,this paper will research and optimize the super-resolution image reconstruction algorithm,image denoising algorithm and image warping correction algorithm based on deep learning in video pre-processing algorithms.The main contents are as follows:(1)Research on super-resolution image reconstruction algorithm.This thesis proposes a clique network super-resolution image reconstruction algorithm based on Laplace pyramid structure.The algorithm uses a Laplacian pyramid structure to gradually reconstruct high resolution images through feature extraction and image reconstruction.Use improved group block as the building block of the network.There are both forward and feedback connections between convolutional layers,and the information between layers is updated alternately.So that the information flow and feedback mechanism can be maximized,the connections between layers are more dense.At the same time,residual learning is used in the network to reduce network parameters and avoid gradient explosions.(2)Research on image denoising algorithm.This thesis designs a fast image denoising algorithm based on OCTNet.The algorithm uses the Oct Conv convolution structure as part of the basic building blocks of the network.The building block decomposes the input into high-frequency and low-frequency feature maps,and the information update in the convolutional layer uses intra-frequency feature updates and inter-frequency feature updates,thereby effectively reducing the consumption of network computing resources.After that,the batch normalization and activation functions are performed to obtain the output of the building blocks.At the same time,residual learning is used in the network,so we can train residual mapping through residual learning.The combination of residual learning and batch normalization not only accelerates the convergence speed during network training,but also avoids the problems caused by vanishing gradients and gradient explosions.(3)Research on image warping correction algorithm.This thesis studies and optimizes existing image warping correction algorithms.The network structure of the algorithm consists of a shape network and a texture mapping network.The shape network performs 3D reconstruction of warped images,and the texture mapping network maps 3D models to 2D planes,thereby completing warping correction of the images.In this thesis,the U-Net ++ network structure is designed to optimize the shape network.This structure can integrate the characteristics of different frequencies by means of feature superposition,and make full use of the effective features of different depths to improve the overall performance of the network.At the same time,the optimization of the network using pruning is implemented to reduce network parameters.