![Advanced 2D Otsu Method
Xin Li(xl553)*, Jingyao Ren(jr986)**
* Department of Biomedical Engineering, Cornell University
** Department of Electrical and Computer Engineering, Cornell University
Introduction
References
Description of 2D Otsu Method
Experiment Results of Shadow Image
Experiment Results of Noise Image
Final Project
Group 13
ECE 5470
Fall 2015
In this project, we implemented the traditional Otsu method[1] for
image thresholding. Next, based on the weaknesses of the
traditional Otsu method such as sensitive to noise and shadow, we
implemented an Advanced 2D Otsu method[2,3]. Different tests
were conducted to evaluate the performance of the Advanced 2D
Otsu method. Comparing to traditional Otsu, the 2D Otsu method is
a better automatic thresholding method for noise and shadow
images, and it could get more accurate thresholding than the
traditional Otsu method.
The keypoint of the 2D Otsu method is to use both the distribution
of the pixel grayscales and the spatial information to decide the
threshold together[2,3]. The 1D Otsu method only consider the
gray level of the image and ignored the spatial information. To
overcome this problem, we could use 2D variance-based
techniques using local neighborhood as well as pixel information
by maximizing the between-class variance defined on the 2D
histogram. This will improve the noise robustness of 1D Otsu.
Compared to traditional Otsu method, 2D Otsu shows better
performance at eliminating shadows and background noise.
However, compared to traditional Otsu with Mean Filter, 2D Otsu
method shows a similiar performance and limited improvements
on denoising. In conlcusion, 2D Otsu is a good choice for
removing shadows but not a perfect choice for denoising. Future
work are needed to improve its denoising ability of 2D Otsu
method.
[1]. Nobuyuki Otsu (1979). "A threshold selection method from gray-level histograms". IEEE Trans. Sys., Man., Cyber. 9 (1): 62–66
[2]. Jianzhuang, Liu, Li Wenqing, and Tian Yupeng. "Automatic thresholding of gray-level pictures using two-dimension Otsu method."
Circuits and Systems, 1991. Conference Proceedings, China., 1991 International Conference on. IEEE, 1991.
[3]. Nie, F., Wang, Y., Pan, M., Peng, G., & Zhang, P. (2013). Two-dimensional extension of variance-based thresholding for image
segmentation. Multidimensional systems and signal processing, 24(3), 485-501.
[4]. Berkeley Segmentation Dataset and Benchmarks 500 (BSDS500), website source: http://www.eecs.berkeley.edu/Research/Projects/
CS/vision/grouping/resources.html
1
2
4
5
2D histogram 1D projection of 2D
histogram
MSE functions of 2D Otsu. Maximizing
Z* will achieve best threshold
1D projection of 2D histogram
Experiment Design3
Design:
• Dataset: Use images with noise(i.e. Gaussain noise, Salt and Pepper
noise, etc. ) and shadow from BSD500 database[4].
• Ground True: Manually mark the ground truth.
• Evaluation: Jaccard and Dice Index; compare result image directly
Additional test:
• Compare the thresholding performance of 2D Otsu on noise image
and original Otsu on Mean-Filtered image.
Hypothesis:
• The 2D Otsu method has better thresholding performance than
original Otsu espcially on images with noise and shadow.
Fish Otsu 2D Otsu
Mp Otsu 2D Otsu
Beaver Otsu 2D Otsu
Image Name
Jaccard
Index
Dice
Index
Normalized Jaccard Index
(2D Otsu is 1) Improvement
(Jaccard)
Otsu 2D Otsu Otsu 2D Otsu Otsu 2D Otsu
Fish 0.01542 0.07785 0.03038 0.14446 0.5075 1 49.24%
Mp 0.05208 0.17051 0.09901 0.29135 0.5260 1 47.40%
Beaver 0.19720 0.33464 0.32943 0.50146 0.5986 1 40.14%
Cup 0.06186 0.10345 0.11651 0.18750 0.5309 1 46.91%
Cup Otsu 2D Otsu
Rectangle Rectangle + Gaussain noise
Otsu only 2D Otsu Mean Filter + Otsu
The 2D Otsu method could work as
a Low-Pass-Filter to remove noise.
To evaluate its performance, we
also compared the thresholding
quality of the 2D Otsu with Mean
Filter + Otsu.
Conclusion6
Eagle + Salt and Pepper noise Otsu only
Mean-Filter + Otsu2D Otsu](https://image.slidesharecdn.com/cvposter-160222075432/85/Advanced-2D-Otsu-Method-1-320.jpg)
Advanced 2D Otsu Method
- 1. Advanced 2D Otsu Method Xin Li(xl553)*, Jingyao Ren(jr986)** * Department of Biomedical Engineering, Cornell University ** Department of Electrical and Computer Engineering, Cornell University Introduction References Description of 2D Otsu Method Experiment Results of Shadow Image Experiment Results of Noise Image Final Project Group 13 ECE 5470 Fall 2015 In this project, we implemented the traditional Otsu method[1] for image thresholding. Next, based on the weaknesses of the traditional Otsu method such as sensitive to noise and shadow, we implemented an Advanced 2D Otsu method[2,3]. Different tests were conducted to evaluate the performance of the Advanced 2D Otsu method. Comparing to traditional Otsu, the 2D Otsu method is a better automatic thresholding method for noise and shadow images, and it could get more accurate thresholding than the traditional Otsu method. The keypoint of the 2D Otsu method is to use both the distribution of the pixel grayscales and the spatial information to decide the threshold together[2,3]. The 1D Otsu method only consider the gray level of the image and ignored the spatial information. To overcome this problem, we could use 2D variance-based techniques using local neighborhood as well as pixel information by maximizing the between-class variance defined on the 2D histogram. This will improve the noise robustness of 1D Otsu. Compared to traditional Otsu method, 2D Otsu shows better performance at eliminating shadows and background noise. However, compared to traditional Otsu with Mean Filter, 2D Otsu method shows a similiar performance and limited improvements on denoising. In conlcusion, 2D Otsu is a good choice for removing shadows but not a perfect choice for denoising. Future work are needed to improve its denoising ability of 2D Otsu method. [1]. Nobuyuki Otsu (1979). "A threshold selection method from gray-level histograms". IEEE Trans. Sys., Man., Cyber. 9 (1): 62–66 [2]. Jianzhuang, Liu, Li Wenqing, and Tian Yupeng. "Automatic thresholding of gray-level pictures using two-dimension Otsu method." Circuits and Systems, 1991. Conference Proceedings, China., 1991 International Conference on. IEEE, 1991. [3]. Nie, F., Wang, Y., Pan, M., Peng, G., & Zhang, P. (2013). Two-dimensional extension of variance-based thresholding for image segmentation. Multidimensional systems and signal processing, 24(3), 485-501. [4]. Berkeley Segmentation Dataset and Benchmarks 500 (BSDS500), website source: http://www.eecs.berkeley.edu/Research/Projects/ CS/vision/grouping/resources.html 1 2 4 5 2D histogram 1D projection of 2D histogram MSE functions of 2D Otsu. Maximizing Z* will achieve best threshold 1D projection of 2D histogram Experiment Design3 Design: • Dataset: Use images with noise(i.e. Gaussain noise, Salt and Pepper noise, etc. ) and shadow from BSD500 database[4]. • Ground True: Manually mark the ground truth. • Evaluation: Jaccard and Dice Index; compare result image directly Additional test: • Compare the thresholding performance of 2D Otsu on noise image and original Otsu on Mean-Filtered image. Hypothesis: • The 2D Otsu method has better thresholding performance than original Otsu espcially on images with noise and shadow. Fish Otsu 2D Otsu Mp Otsu 2D Otsu Beaver Otsu 2D Otsu Image Name Jaccard Index Dice Index Normalized Jaccard Index (2D Otsu is 1) Improvement (Jaccard) Otsu 2D Otsu Otsu 2D Otsu Otsu 2D Otsu Fish 0.01542 0.07785 0.03038 0.14446 0.5075 1 49.24% Mp 0.05208 0.17051 0.09901 0.29135 0.5260 1 47.40% Beaver 0.19720 0.33464 0.32943 0.50146 0.5986 1 40.14% Cup 0.06186 0.10345 0.11651 0.18750 0.5309 1 46.91% Cup Otsu 2D Otsu Rectangle Rectangle + Gaussain noise Otsu only 2D Otsu Mean Filter + Otsu The 2D Otsu method could work as a Low-Pass-Filter to remove noise. To evaluate its performance, we also compared the thresholding quality of the 2D Otsu with Mean Filter + Otsu. Conclusion6 Eagle + Salt and Pepper noise Otsu only Mean-Filter + Otsu2D Otsu