Registration Methods for VNREDSat-1 Images

My thesis

Transcript

1. Registration Methods for VNREDSat-1 Images Nguyen Thac Du K55CA
2. Outline VNREDSat-1 satellite Problem statement Problem solution Result Conclusion 5/27/2014 2
3. VNREDSat-1 satellite Launched in May 2013 Monitor and Study Not much research related to its image datas 5/27/2014 3
4. VNREDSat-1 Images Panchromatic Multispectral 5/27/2014 4
5. VNREDSat-1 Images Panchromatic Multispectral Captured at the same scene 5/27/2014 5 2.5m 10m
6. Registration The Ground PAN region MS region geometric error  Image registration: register MS to the same coordinate of PAN 5/27/2014 6
7. Proposed Solutions Based on Block Matching Algorithm SIFT and Interpolation 5/27/2014 7 *SIFT: Scale-invariant feature transform
8. Block Matching Computer Vision methods Determine matching blocks Estimate motion of each pixel Output: disparity vector matrix 5/27/2014 8
9. Block Matching 5/27/2014 9 = 1 2 =0 −1 =0 −1 − Use a cost function Least cost means closest matching Target Reference Disparity Vector Macro Block Best Match
10. 5/27/2014 10 Method based on BMA
11. SIFT Computer Vision methods Detects and describes features in image Features are distinctive and invariant to transform 5/27/2014 11
12. 5/27/2014 12 Method based on SIFT
13. 5/27/2014 13 Divide Match and Construct Disparity Vector Matrix Upsample
14. Result 5/27/2014 14 Size of MS Implementation Time BMA SIFT 500x500 pixels 79.36 6.78 1000x1000 pixels 315.47 16.20 1750x1750 pixels 978.89 49.17
15. 5/27/2014 15 Original MS BMA SIFT
16. 5/27/2014 16 Original MS BMA SIFT
17. Result 5/27/2014 17 Experiment MAD BMA SIFT 1 2.930 11.338 2 3.975 16.580 = 1 2 =0 −1 =0 −1 −
18. Conclusion BMA High accuracy Slow implementation Color is distorted SIFT Not corresponding in some part Fast Color is almost preserved 5/27/2014 18
19. Thank You For Listening!