Training-based Super Resolution Enhancement using CUDA D 張書豪 R 張嫚家 R 楊逸民
2 Outline Introduction Training-based super resolution method Improvement and application Task and Goal Schedule
3 Introduction Super resolution can enhance the resolution of images Image upsampling
4 Introduction Related work thumbnail image low resolution image -> high resolution image Make better quality surveillance system online video quality
5 Introduction Motivation SR is useful in generating high resolution images Yet very time-consuming
6 Original method Background Hong Chang, Dit-Yan Yeung, Yimin Xiong, “Super Resolution through Neighbor Embedding,” CVPR 2004 Training-based super resolution Multiple training patches from different images Preserve the low resolution and high resolution correspondence
7 Flowchart
8 Improvement Pre-processing compute energy map to separate low and high frequency region LR patches can be parallelly processed compute LR feature vector K-NN search weight vector HR patch composition
9 Tasks and Goals Get high resolution images in short time Extend super resolution to video sequences Example youtube video sequences Fast loading on high resolution video such as 480p, 720p
10 Schedule In project midterm Finish image super resolution on CUDA one month later In project final Finish video super resolution on CUDA
11 Video version