![21
Simulation results
Comparison between our algorithm and
traditional level set methods proposed in [1
and 2]](https://image.slidesharecdn.com/icsipatexture-170617190037/85/Locating-texture-boundaries-using-a-fast-unsupervised-approach-based-on-clustering-algorithms-fusion-and-level-set-21-320.jpg)
![27
Summary
In this paper, a fast level set based method has been
proposed for image segmentation.
Our algorithm is robust against noise.
The proposed feature space has much less
dimensionality compared to Gabor and structure
tensors.
Unlike [1], image gradients have not been calculated,
which decreases the effects of noise.
Using fusion, significantly increases the
generalization of the clustering algorithms.](https://image.slidesharecdn.com/icsipatexture-170617190037/85/Locating-texture-boundaries-using-a-fast-unsupervised-approach-based-on-clustering-algorithms-fusion-and-level-set-27-320.jpg)
![28
References
[1] S. Daniel Cremers, M. Rousson, and R. Deriche, "A Review of Statistical Approaches
to level sets Segmentation: Integrating Colour, Texture, Motion and Shape", 2007,
International Journal of Computer Vision 72(2), 195–215
[2] M. Rousson, T. Brox, and R. Deriche, "Active Unsupervised Texture Segmentation on
a Diffusion Based Feature Space", 2003, Proceedings of the 2003 IEEE Computer
Society Conference on Computer Vision and Pattern Recognition (CVPR’03)](https://image.slidesharecdn.com/icsipatexture-170617190037/85/Locating-texture-boundaries-using-a-fast-unsupervised-approach-based-on-clustering-algorithms-fusion-and-level-set-28-320.jpg)
Locating texture boundaries using a fast unsupervised approach based on clustering algorithms fusion and level set
- 1. 1 Locating Texture Boundaries Using a Fast Unsupervised Approach Based on Clustering Algorithms Fusion and Level Set Slides by: Mehryar Emambakhsh Sahand University of Technology
- 2. 2 Outline About image segmentation and its methods Feature extraction – Color transformation – Non-linear diffusion Clustering algorithms – Fusion Level set Simulation results Summary References
- 3. 3 About image segmentation and its methods Image segmentation is a procedure in which an image is partitioned into its constituting regions. There must be a uniformity in some predefined features in each region: – Pixels intensity – Color components – Texture features – Motion vectors
- 4. 4 About image segmentation and its methods There are many different approaches for image segmentation: 1) Clustering-based methods feature space clustering algorithms Advantage: – Fast computational speed Disadvantages: – Sensitivity to noise and outliers in the feature space – Over-segmentation
- 5. 5 About image segmentation and its methods 2) Energy minimization methods feature space energy function minimization Advantages: – Reasonable results – Robust against noise Disadvantages: – High computational complexity – Sensitive to local minima
- 6. 6 About image segmentation and its methods Our proposed method:
- 7. 7 Feature extraction: color transformation Color transformation: – Non-linear color spaces generate a more separable feature space compared to linear color spaces. – Among non-linear color spaces, CIE L*a*b*, which is a uniform color space, produces a much detachable feature space compared to the non- uniform ones.
- 8. 8 Feature extraction: non-linear diffusion Non-linear diffusion is a method for image de-noising and simplification. It is used for feature extraction from texture in our approach. Non-linear diffusion equation is solved on the color image: g(.) is a decreasing function of image gradient. Non-linear diffusion has many superiority compared to other texture feature extraction methods: – Low dimensionality – Preserving image edges – Robust against noise
- 9. 9 Clustering algorithms Fuzzy C-means (FCM), K-means, SOM (Self- Organizing Map), and GMM (Gaussian Mixture Model) have been evaluated in our work. FCM is a clustering technique wherein each data point belongs to a cluster to some degree that is specified by a membership degree. However, K-means assigns each point to the cluster whose center (centroid) is nearest. – Euclidean distance is used in our work because of its better performance than city-block and Hamming distance criteria. – Also it is faster that Mahalanobis distance.
- 10. 10 Clustering algorithms The other clustering algorithm that we have utilized is SOM neural network. It is an unsupervised competitive neural network. The structure of the neural network is as follows:
- 11. 11 Clustering algorithms Finally, GMM is our last clustering algorithm. In GMM, each mass of features is modeled as multivariate normal density function. These models are fit to data using expectation maximization algorithm, which assign a posteriori probability to each observation. The dependency of each pixel to a specific cluster is determined by examining the value of the probability.
- 12. 12 Clustering algorithms: fusion Choosing a clustering method depends on the input data distribution. – Highly overlapped feature space SOM – Moderately overlapped feature space FCM and K-means – Feature space with suitable detachability GMM To incorporate these clustering algorithms, a fusion of them is used here.
- 13. 13 Level set Unlike previous algorithms, the cluster map is used to evolve the contour. This significantly, decreases the computational complexity.
- 14. 14 Simulation results Our algorithm has been evaluated on an Intel Core 2 Due CPU (T7250). 59 images from Corel texture dataset has been used. The average values for : 120 and 80 epochs for training the first and the second SOM stages, respectively. 1543.0,2462.0,3176.0,2819.0 4321 ==== αααα iα
- 15. 15 Simulation results The input image and the ground truth Color transformation result Non-linear diffusion result
- 16. 16 Simulation results The clustering results
- 17. 17 Simulation results (a) contour initialization (b) 60th iteration (c) 100th iteration (d) The final segmentation result
- 18. 18 Simulation results PSNR = 28.22 (db) (a) contour initialization (b) 20th iteration (c) 120th iteration (d) The final segmentation result
- 19. 19 Simulation results PSNR = 27.63 (db) (a) contour initialization (b) 60th iteration (c) 120th iteration (d) The final segmentation result
- 20. 20 Simulation results Clustering algorithms performance: PSNR vs. PCS
- 21. 21 Simulation results Comparison between our algorithm and traditional level set methods proposed in [1 and 2]
- 23. 23 Simulation results PSNR = 28.53 (db) (a) The input image (b) initialization (c) 60th iteration (d) The final segmentation result
- 24. 24 Simulation results PSNR = 26.72 (db) (a) The input image (b) initialization (c) 60th iteration (d) The final segmentation result
- 25. 25 Simulation results We claimed that CIE Lab color space is highly more suitable for setting up a feature space, instead of RGB color space.
- 27. 27 Summary In this paper, a fast level set based method has been proposed for image segmentation. Our algorithm is robust against noise. The proposed feature space has much less dimensionality compared to Gabor and structure tensors. Unlike [1], image gradients have not been calculated, which decreases the effects of noise. Using fusion, significantly increases the generalization of the clustering algorithms.
- 28. 28 References [1] S. Daniel Cremers, M. Rousson, and R. Deriche, "A Review of Statistical Approaches to level sets Segmentation: Integrating Colour, Texture, Motion and Shape", 2007, International Journal of Computer Vision 72(2), 195–215 [2] M. Rousson, T. Brox, and R. Deriche, "Active Unsupervised Texture Segmentation on a Diffusion Based Feature Space", 2003, Proceedings of the 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’03)
- 29. 29 Thanks for your attention!