1. 1 Iterative Multimodel Subimage Binarization for Handwritten Character Segmentation Author: Amer Dawoud and Mohamed S. Kamel Source: IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO. 9, SEPTEMBER 2004, pp. 1223-1230 Speaker: Ching-Hao Lai( 賴璟皓 ) Date: 2004/10/13

2. 2 Outline Introduction Iterative Multimodel Binarization Algorithm Feature Extraction Setting Rejection Criteria Selecting Subimages Optimal Thresholds Experimental Results Conclusion

3. 3 Introduction(1/3) Existing binarization methods: Global binarization method Local binarization method A document image is divided into subimages: Image(1) Image(2) … Image(M). To find an optimal threshold for each subimage that would eliminate background noise.

4. 4 Introduction(2/3)

5. 5

6. 6 Introduction(3/3) The proposed method uses multimodels to iteratively arrive at the optimal threshold for each subimage. Based on gray-level and stroke-run

7. 7 ITERATIVE MULTIMODEL BINARIZATION ALGORITHM The subimages are then binarized at a sequence of candidate thresholds CT i, where CT 1 is the lowest possible threshold in gray-scale histogram. The difference between two successive CTs was chosen to be eight gray-levels, which we found to be satisfactory.

8. 8 Feature Extraction When CTi failed to eliminate the background noise in Image (x). We want to infer such failure by comparing features of the binarized Image (x) with those of the other subimages. Features: Gray-Level Features Stroke-run Features

9. 9 Gray-Level Feature

10. 10 Gray-Level Feature

11. 11 Stroke-Run Feature Stroke-Run historgram: K={1,2,…,M}, 4200 images, the longest run is 5 pixels Unit-Run:

12. 12 Setting Rejection Criteria Stroke-Width feature: GRC: gray-level rejection criterion, SRC: stroke rejection criterion

13. 13 Flowchart

14. 14

15. 15 Comparison result

16. 16 Comparison result

17. 17 Conclusions When applied to a set of images that represent wide range of background complexity and noise levels, the multimodel algorithm succeeded in eliminating the background, and in preserving the handwritten characters. As a result, higher recognition rate and lower substitution, insertion, and deletion error rates were achieved.