Research And Application Based On Video Super Resolution

Super-resolution technology is a process of obtaining clear high-resolution image by processing a series of low-resolution images.Its essence is to recover high-resolution video image by using the spatial information redundancy of the image itself and the time information redundancy between adjacent frames in the video frame,so as to obtain more delicate image quality and improve the visual experience of the human eye.In the process of video image transmission,in order to save bandwidth,high-definition video is often compressed and encoded first and transmitted to the receiver for decoding,but the recovery after compression is lossy,and the recovered image will lose part of the information.Using super-resolution technology as the post-processing algorithm of video image can improve the video quality and improve the appearance.At present,the research of video super-resolution technology has made rapid progress,but how to fully integrate spatiotemporal information and how to restore real and natural texture still needs to be further studied.This thesis mainly studies how to improve the visual quality and get a better view on the basis of 4x resolution restoration of daily video.Firstly,for the preprocessing module,semantic segmentation network is introduced to provide more prior texture information for the preprocessing module.Different objects have different textures,but if the colors of adjacent objects are very similar,the super-resolution network may recognize them as one kind,which leads to the unreal texture recovery.Therefore,this thesis introduces the prior information of semantic segmentation,which makes the super-resolution network recover more real texture,and improves the super-resolution effect by 0.59 dB.Secondly,for the fusion module,three-dimensional fusion technology is introduced.Although the common two-dimensional convolution can get different information from each layer's characteristic graph,the information of each layer's convolution kernel is not connected,and it can't well fuse the timing information of video itself.After adding three-dimensional fusion technology,the model can better fuse video timing information,and improve the super-resolution effect by 0.17 dB.Finally,for the reconstruction module,the channel attention mechanism is introduced.Although in the process of feature extraction,the feature map can be very deep,but not every layer of feature map is the same important.After the introduction of channel attention mechanism,more attention can be paid to the layer that contributes more to the results,so as to get better results.Aiming at the problem of insufficient information in the general average pooled attention channel,the maximum pooled attention channel is introduced to supplement the information,and the video super-resolution effect is improved by 0.42 dB.At the end of this thesis,the experimental results show that the algorithm in this thesis is better than the previous ones.