| Video super-resolution(VSR)reconstruction technology aims to improve the spatio-temporal resolution of continuous low-resolution(LR)frames.It has very important research value in the fields of medicine,military,astronomy,surveillance and security,and has become a hot research direction in machine vision.In the industrial field,reconstructing video under low-light conditions to obtain clear imaging results has received more and more attention,it is of great significance for the optimization of production line.Therefore,in this paper,three different video super resolution reconstruction models are proposed to improve the accuracy of video reconstruction and reduce the complexity of the model based on the dim industrial environment and considering the problems of feature extraction and model operation cost.The main work of this paper is as follows:(1)Aiming at the accuracy of detail information extraction of each frame in the video,in order to better utilize the complementary information of adjacent frames of the video,and realize the fusion of LR frame detail features into highresolution(HR)frames,this study designed a fusion block,which uses the hidden information generated by multi-frame to fuse with the corresponding LR frames,accurately extracts the associated features of each frame.The experiment show that the network proposed has better results than some other network structures in the evaluation indicators PSNR and SSIM.(2)In order to make full use of the representation learning ability of convolutional neural network and strengthen the attention on the correlation of pixel features within each frame of video,this paper proposes a dual attention module.Based on the network model in Chapter 2,this module is added when extracting the features of a single frame.This module does not add any parameters to the overall network,and realizes the prediction of feature channels and spatial pixels and their weights.Finally,comparing the model in this chapter with other attention mechanisms,the results show that our module has better performance in the VSR reconstruction network.It is found that the proposed model has higher reconstruction accuracy by comparing with the second chapter model and other VSR models(3)In view of the high hardware requirements and large number of model parameters of the above network model,the operation cost of video reconstruction is relatively high.In order to reduce the number of parameters and calculations of CNNs,this study improves the residual block in Chapters 2 and 3,and introduces the idea of multi-scale feature extraction.The Ghost module is used to process redundant features to form a lightweight progressive fusion block to reduce the number of parameter.The experimental results show that compared with the model PDAVSR of Chapter 3,the number of parameter and floating point operation of the model in this chapter are reduced by 30%,the value of PSNR is reduced by nearly 1%,and the value of SSIM is reduced by 0.01%.On the premise of having fewer parameters,the proposed model in this study has better performance and lower model complexity. |
