Presentation is loading. Please wait.

Presentation is loading. Please wait.

Accelerating a Satellite Image Processing using GPU Jihoon Kang, In-Hoi Koo, and Sang-Il Ahn Ground System Development Department Korea Aerospace Research.

Similar presentations


Presentation on theme: "Accelerating a Satellite Image Processing using GPU Jihoon Kang, In-Hoi Koo, and Sang-Il Ahn Ground System Development Department Korea Aerospace Research."— Presentation transcript:

1 Accelerating a Satellite Image Processing using GPU Jihoon Kang, In-Hoi Koo, and Sang-Il Ahn Ground System Development Department Korea Aerospace Research Institute

2 Why High Performance Computing? High Resolution Earth Observation Satellite –KOMPSAT-2 1m Resolution 15,000 pixel x 15,000 pixel –KOMPSAT-3 0.7m Resolution 24,000 pixel x 24,000 pixel –Data size increased 2.56 times –About 100 Minutes Contact Interval High Resolution Geostationary Satellite –COMS (MI) 1km Resolution Visible Band, 5 Spectral Band Available –GEOKOMPSAT-2A (planed) 0.5 km Resolution Visible Band, more than 14 Spectral Band –Data size increased about 30 times –Data transmitted to Ground Station during 24/7 Require FASTER image processing method

3 Image Processing Algorithm using GPU Good –Pixel by Pixel wise operation –Simple arithmetic operation –Large Size of Input Data But GPU Memory Capability is limited Bad –Individual output pixel depends on previous result –Complex logical flow –Small Size of Input Data

4 Satellite Image Processing Decompression Non-Uniformity Correction Filtering MTF Compensation Interpolation Registration FFT & IFFT Convolution like operation can be good candidates for using GPU –Spatial domain based Filtering, MTFC, and Interpolation And, output data independent –Non-Uniformity Correction

5 Cubic Convolution Interpolation Convolution like operation –Simple Arithmetic Operation Included Output Independent Each thread of GPU calculates one output pixel 4x4 input matrix is stretched

6 Result

7

8 MTF Compensation 7x7 MTF Kernel Convolution Operation Includes Only Addition and Multiplication Each thread of GPU calculates one output pixel

9 Result

10

11 Analysis GPU outperforms CPU –Even GPU based code is not optimized –Cubic Convolution Interpolation About 40 times faster –MTF Compensation About 28 times faster If Target Data Size is Small, GPU takes more time to process –Memory allocation GPU –Data copy time to GPU –Result data copy time to Main Memory –Thread Management Overhead

12 Conclusion Satellite Image Data is getting bigger –Image Processing system is required to be FASTER GPU based image processing is FAST –But, target, which will be run on GPU, should be carefully selected and re-coded –Reduce data copy to GPU and bring back to Main Memory time is more efficient In Future, –Process all possible image processing algorithm in GPU –Now, working on Wavelet based Decompression

13 Appendix A – Computing Environment CPU: Intel Core 2 Extreme QX9650, 3GHz Memory: DDR2-8GB GPU: NVIDIA Tesla C2050 Software –Windows 7 Enterprise K 64-bit –MS Visual Studio 2010, C++ –NVIDIA CUDA 4.0

14 Appendix B – Measured Time Table MTF Compensation (in msec) Input Data Size7x7256x256512x x x x x x16384 CPU (1 Thread) CPU (4 Thread) GPU Cubic Convolution Interpolation (in msec) Input Data Size7x7256x256512x x x x x x16384 CPU (1 Thread) CPU (4 Thread) GPU


Download ppt "Accelerating a Satellite Image Processing using GPU Jihoon Kang, In-Hoi Koo, and Sang-Il Ahn Ground System Development Department Korea Aerospace Research."

Similar presentations


Ads by Google