11
Super-resolution on Video with User Interface 薛健鑫, 蔡昆修, 何星翰 1

Gpgpu presentation 2

Embed Size (px)

DESCRIPTION

The check point presentation of gpu project.

Citation preview

Page 2: Gpgpu presentation 2

Super-resolution on Video

Our goal : User can zoom-in the video with high resolution

result when watching video.

2

Page 3: Gpgpu presentation 2

Project Goals (Original)

Retrieving image stream

From raw image sequence

From DirectX

Super-resolution processing

Algorithm survey

GPU speedup investigation

Evaluation method

Graphic User Interface

Video player

Zoom in/out

3

Page 4: Gpgpu presentation 2

Current Status

Retrieving image stream

From raw image sequence

From OpenCV

Super-resolution processing

Algorithm survey

GPU speedup investigation

Evaluation method

Graphic User Interface

Video player

Zoom in/out

4

Page 5: Gpgpu presentation 2

Result Demo

5

Page 6: Gpgpu presentation 2

THE END

6

Page 7: Gpgpu presentation 2

Additional Material

7

Page 8: Gpgpu presentation 2

About SR algorithm Reference: Practical Super-Resolution from Dynamic Video Sequences

Input: A target LR image and a sequence of neighbor LR image

Step 1. get initial HR image f0

Step 2. use forward projection to simulate those LR image.

Step 3. use the difference between ground truth LR and simulate

LR to do backward projection and make guess HR image f1

Step 4. After n iteration we get a HR image fn which can

simulate LR images which are very closed to ground truth.

Wei-Chao Chen

([email protected])8

Page 9: Gpgpu presentation 2

Some Problems about IBPWe know the flow of algorithm, but some detail problems exist

1. How to get good f0 ?

2. The optical flow relation between LR neighbors and target

HR image.

3. The ‘Quality Pair’ examination algorithm in the paper ,

which is not clear in the paper.

4. BP algorithm with EWA filtering

Those questions are now bothering us and we need further

studies to solve them.

Wei-Chao Chen

([email protected])9

Page 10: Gpgpu presentation 2

The initial estimation of HR

We regard the IBP as a further refinement , base on a good

initial upscaling result. It is also important to have a very

good f0 estimation in the IBP algorithm.

Bicubic resizing has very good result in upscaling and easy

to implement, so it is our first choice.

Wei-Chao Chen

([email protected])10

Page 11: Gpgpu presentation 2

Retrieving image stream from video

Extract video frames by OpenCV

Load frame sequence in OpenGL as textures

Display video using texture animation technique

Wei-Chao Chen

([email protected])11