Fast CCL(connected component labeling) with GPU
1 like2,346 views
This document summarizes a parallel approach to object identification in large-scale images presented at ICESS 2016 in Takamatsu, Japan. The approach exploits data parallelism on the GPU for connected component labeling (CCL) in three main steps: initializing pixel labels, computing column-wise label runs, and efficiently merging labels through parallel boundary comparisons. Experimental results on images up to 4096x4096 pixels with varying object densities and shapes demonstrate the efficiency of the GPU approach compared to a reference CPU implementation from OpenCV. The method can be successfully applied to applications involving large-scale images like object tracking with radar signals.
Fast CCL(connected component labeling) with GPU
- 1. ICESS 2016, Takamatsu, Japan 14 ~ 16 Nov. 2016 Young-Min Kang Tongmyong University A Parallel Approach to Object Identification In Large-scale Images Sung-Soo Kim, ETRI Gyung-Tae Nam, GCSC Inc.
- 2. Bigger images • Era of Big data – Increased sizes of images data • Image processing – Heavy Computation • One of the most fundamental operations – Object identification/recognition • Image segmentation • Connected components labeling
- 3. Connected component labeling • Objective – Pixels in a connected component have an identical labels
- 4. Parallel image processing • Most image processing algorithms – Pixel-wise operations • can be implemented with pixel-wise threads • can be efficiently performed in a data-parallel fashion • GPU – Data parallel device – can be easily applied to various image processing methods GPU: Many-core architecture
- 5. Pixel connectivity • Graph representation Image Pixel connectivity
- 6. CCL and parallelism • CCL with graph traversal – cannot be easily parallelized • Traversal = sequential • GPU based approaches – has not been very successful
- 7. Our method • GPU-based efficient algorithm for CCL – Data initialization – Computing column-wise label runs – Efficient label merge
- 8. Data initialization • Each pixel is assigned unique label if it is turned on
- 9. Data initialization • Each pixel is assigned unique label if it is turned on 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
- 10. Data initialization • Each pixel is assigned unique label if it is turned on 1 2 -1 -1 -1 6 7 -1 9 10 11 12 -1 14 -1 -1 -1 -1 19 20 -1 -1 -1 -1 -1
- 11. Column-wise label runs • Run – Block of contiguous object pixels in a column • Computing column-wise label runs – Can be done with w threads h w
- 12. Column-wise label runs • Label change within a column (1 thread)
- 13. Column-wise label runs • Graph-based interpretation
- 14. Column-wise label runs • Implementation
- 15. Label merge • After computing “column-wise label runs” – We have separate trees to be merges in accordance with their connectivity • What is needed – Checking vertical adjacency
- 16. Label merge • Connectivity check
- 17. Label merge • Connectivity check
- 18. Label merge • Connectivity check
- 19. Label merge • Connectivity check
- 20. Label merge • Updated hierarchy
- 21. Why only roots are changed Let’s merge OK! I will follow you
- 22. Why only roots are changed Merged tree
- 23. Previous methods 1. Check the connectivity 2. Update the hierarchy 3. Iterate this process until no update is made A kind of graph traversal Heavy computation when the pixels make a long connected chain
- 24. Our method • Label merge is performed with fixed number of iterations – The number of iteration • log2(w) – Computation cost at every iteration • reduced to be the half the previous one • Efficient label merge • Moreover – Can be easily parallelized
- 25. Label merge boundary • 1st merge w/2 boundaries h comparisons in each boundary wh/2 threads
- 26. Label merge boundary • 2nd merge w/22 boundaries h comparisons in each boundary wh/22 threads
- 27. Label merge boundary • 3rd merge w/23 boundaries h comparisons in each boundary wh/23 threads
- 28. Label merge boundary • Final merge log2(w) –th merge Computation cost at the 1st merge: C(1) Total Cost
- 30. Performance • Computational cost for each task – Cost for Initialization = 1 – 4096x4096 images with different number of connected components 50 labels 1869 labels initialization 1.0 1.0 column-wise run 1.6 1.6 label merge 3.4 3.6
- 31. Performance • Computational cost for each task – Cost for Initialization = 1 – 4096x4096 images with different number of connected components 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 initialization column-wise run label merge 50 labels 1869 labels
- 32. Experimental results • Reference – Grana’s method implemented with OpenCV • Two Tests – Random noise with varying densities – Object identification with shapes
- 33. Varying densities • Image size: 2048x2048
- 34. Varying densities • Image size: 2048x2048
- 35. Varying densities • Image size: 4096x4096
- 36. Varying densities • Image size: 4096x4096
- 37. Object identification with shapes • Two spiral curves
- 38. Object identification with shapes
- 39. Object identification with shapes • Stars
- 40. Object identification with shapes • Stars
- 41. Applications • Object tracking with radar signal
- 42. Conclusion • An efficient GPGPU implementation for CCL • Data-parallelism of GPU exploited • Experimental results show its efficiency • Can be successfully applied to various applications with large-scale images – e.g., Object identification from radar signals