多尺度可變形卷積對齊網路應用於影片超解析;Multiscale Deformable Convolution Alignment Network for Video Super Resolution

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89997

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89997

Title:	多尺度可變形卷積對齊網路應用於影片超解析;Multiscale Deformable Convolution Alignment Network for Video Super Resolution
Authors:	李易翰;Lee, Yi-Han
Contributors:	資訊工程學系
Keywords:	影片超解析;可變形卷積;多尺度模型;注意力機制;Video super resolution;Multiscale model;Deformable Convolution;Attention mechanism
Date:	2022-08-30
Issue Date:	2022-10-04 12:07:10 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著科技進步，能夠拍出高解析度影像的設備以及高解析度的顯示器已經隨手可得，許多過去儲存的影片電子檔的解析度相比現今流通的顯示器顯得很低，這時候就需要用到影片超解析演算法來提升影片的解析度，此外影片超解析也能夠應用在網路傳輸，能夠在傳輸影片之前下降影片解析度，傳輸完畢後再利用影片超解析演算法將影片的解析度還原，進而達到節省流量的效果。本篇論文提出了一個基於多尺度可變形卷積來進行影像對齊的影片超解析演算法，本文利用多尺度模型的概念，使用不同解析度的分支來預測可變形卷積的偏差值，進而增強對齊模組，並且使用SE block來整合特徵傳遞階段產生的影像特徵，幫助模型找出重要的特徵用來重建影像。本論文使用Reds資料集以及Vimeo-90k對模型進行訓練以及測試，在Reds提供的測試資料集Reds4上測試能夠超越basicVSR++0.07dB的PSNR，在視覺方面則是能夠生成出較為清晰、銳利的紋理。;As a result of highly developed technology, high-resolution devices and screens are extremely easy to obtain nowadays. The display problem with distorted image which occurs on the current monitor is due to the low resolution of the traditional video. To reconstruct the low resolution video, the video super-resolution techniques are helpful in quickly generating high-resolution video. Hence, this paper proposes a video super-resolution algorithm adopting multi-scale deformable convolution for image alignment to improve the visual quality of the to generate a video with improved visual quality as our final product. In order to enhance the alignment module, the multi-scale model and branches of different resolutions are utilized to predict the deviation value of deformable convolution. Then, the better quality reconstruction video relies heavily on the SE block architecture which is applied to integrate the image features generated from the feature propagation stage in order to help find the better image features for image alignment module. The results of the experiments utilizing the REDs and Vimeo-90k datasets indeed generate better visual quality and high-resolution videos. The proposed algorithm has achieved the best performance. It is 0.07dB higher than the other three methods in PSNR value, and the comparison and ablation experiment results proved the effectiveness of the proposed algorithm.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	71	View/Open

社群 sharing

Loading...