Skip to content

SRGAN + Flow Loss, a loss function regarding both the superresolution quality and temporal continuity for video superresolution based on SRGAN

Notifications You must be signed in to change notification settings

devanshraj3000/SRGAN-with-Flow-Loss

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SRGAN + Flow Loss

W4731 Computer Vision Final Project: Video Superresolution

In this project, a loss function to improve upon the baseline approach which simply applies image supperresolution to each frame will be proposed. Based on the SRGAN proposed by Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, we propose a novel specific loss function which does not take only one single frame but a bundle of temporally neighbouring frames into the GAN. The model takes a sliding window of several consecutive frames as the input for generator network and outputs a bundle of consecutive SR frames. A perceptual loss function regarding both the single frame superresolution quality and the temporal continuity is proposed and the experimental results will evaluate the quality of superresolved videos using this loss function.

From left to right, they are bicubic method, SRGAN + Flow Loss method, and high resolution video: sr_video1

sr_video2

Approach

A new loss function regarding both the single frame superresolution quality and the temporal continuity: approach

Temporal Continuity

SR flows and HR flows comparison while training: SR flows become more and more similar to HR flows with more epochs trained. flow_training

Superresolution Quality

Superresolution quality comparison: The first three rows show a clip of the video with a businessman, which includes (a) LR frame sequence; (b) SR frame sequence using our loss function; (c) HR frame sequence. The second three rows show a clip of the video of a basket of fruits, which includes (d) LR frame sequence; (e) SR frame sequence using our loss function; (f) HR frame sequence. flow_training

Conclusion

In this project, we propose a novel specific loss function regarding both the superresolution quality and temporal continuity for video superresolution based on SRGAN. Experimental results show that our method keep better temporal continuity. However, the SR frame image quality may drop using our method. The future work would focus in more experiments with bigger datasets and novel architecture of GAN to superresolve videos.

About

SRGAN + Flow Loss, a loss function regarding both the superresolution quality and temporal continuity for video superresolution based on SRGAN

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%