Research On Convolutional Neural Network Based Super-Resolution Technology Of Mixed Resolution Video

Most of the information processed by human comes from the image information obtained by vision.Images play an important role in all aspects of human production and life.Low-resolution images have problems such as blurring,low quality,and poor subjective perception.Super resolution has been proposed and developed to improve the quality of images,and is currently widely used in many fields,such as daily entertainment,public surveillance,satellite and medical image processing.Neural networks are currently developing rapidly and are widely used in image processing tasks.The neural network based super-resolution technology can provide sufficient details for the image to be reconstructed.It has a very good performance by learning a large number of high-low resolution pairs.However,in the single-image super-resolution reconstruction,high-performance network structures have problems such as complex structures and too many parameters,they can only be used in specific cases with huge computing resources.In the video super resolution,which is independent of single image super resolution,the prior information of single image reconstruction algorithm is ignored.And the entire network structure is so huge that lacks flexibility.This paper studied and solved these two issues.This paper studied neural network based single image super-resolution algorithms,and proposed a lightweight neural network that can reconstruct high-quality images with fewer resources.The main research contents include:(1)Investigate how to improve the performance of the residual block on the premise that the amount of parameters and calculations are unchanged.(2)Investigate how to introduce channel attention mechanism and make corresponding improvements.(3)Investigate dense connections and residual connections,and propose a more efficient network connection structure.In the experiment,this paper analyzed he modified residual block and the statistics channel attention module,compared the performance of the proposed network structure with the common connection structure,and verifies the feasibility and effectiveness of each part of the improvement.The comparison with other algorithms also proves the superior performance of the proposed lightweight network.For mixed-resolution video,which consists of a few number of high-resolution frames and a large number of low-resolution frames,this paper proposed a video super-resolution framework that can combine single-image algorithm.The main research focuses include:(1)Research on how to integrate single-image algorithms into mixed-resolution video super-resolution tasks without changing the single-image super-resolution algorithm.(2)Investigate how to efficiently use the information of the high-resolution frames in the mixed-resolution video to compensate the target frames to make them have higher reconstruction quality.Through experiments,this paper verifies the rationality of the proposed framework and information compensation network,and analyzes the flexibility of the proposed framework.The comparison with other algorithms proves that the proposed framework can effectively handle the reconstruction task of mixed-resolution video.