![Discussion
● DBSCAN: clustered only water bodies
● K-mean: loss of data
● Mean shift: overall good result
● N-cut: poorest result (time complexity is
too high)
● Watershed: not good for this dataset
Reference [1]
31](https://image.slidesharecdn.com/charmichokshippt-180908095313/85/Charmi-chokshi-ppt-31-320.jpg)
![References
[1] Xia, Xide, and Brian Kulis. "W-Net: A Deep Model
for Fully Unsupervised Image Segmentation." arXiv
preprint arXiv:1711.08506 (2017).
[2] Shi, Jianbo, and Jitendra Malik. "Normalized cuts
and image segmentation." IEEE Transactions on
pattern analysis and machine intelligence 22.8
(2000): 888-905.
[3] Zhou, Yong-mei, Sheng-yi Jiang, and Mei-lin Yin.
"A region-based image segmentation method with
mean-shift clustering algorithm." Fuzzy Systems and
Knowledge Discovery, 2008. FSKD'08. Fifth
International Conference on. Vol. 2. IEEE, 2008.
42](https://image.slidesharecdn.com/charmichokshippt-180908095313/85/Charmi-chokshi-ppt-42-320.jpg)
Charmi chokshi ppt
- 1. Unsupervised Correction of Training Labels project taken under program Presented by: Charmi Chokshi Final year BTech student of Information and Communication Technology Ahmedabad University Duration: May 31, 2018 till July End Guided by: Mr. Pankaj Bodani (Scientist-SE) Space Application Centre-ISRO
- 2. Outline ● Objectives ● Introduction ● Methodology ● Unsupervised Sementic Segmentation Algorithms ● Discussion ● Region Growing Algorithm ● Conclusion ● Future work ● References 2
- 3. Objective ● To work on the preprocessing (Data Cleaning) step of the image segmentation problem using Deep Learning ● To create accurate Training Labels as the input to Neural Network ● To investigate/compare the use of different unsupervised image segmentation techniques for boundary correction 3
- 5. What is Semantic Segmentation? ● In computer vision, semantic segmentation is the process of partitioning a digital image into multiple meaningful segments ● Semantic segmentation is typically used to locate objects and boundaries in images. ● More precisely, it is the process of assigning a label to every pixel in an image such that pixels with the same label share certain characteristics 5
- 6. Application of Image Segmentation ● Driverless car ● Medical imaging ● Object detection ○ Face detection ○ Pedestrian detection ○ Brake light detection ○ Locate objects in satellite images (roads, forests, crops, water bodies, etc.) ● Recognition Tasks ● Traffic control systems 6
- 7. What is Unsupervised Sementic Segmentation? ● Unsupervised = Learning without the help of teacher! ● No Labeled Training Data available for the model to learn from ● The goal of this unsupervised machine learning technique is to find similarities in the data point and group similar data points together which will give us insight into underlying patterns of different groups ● “Clustering” is the process of grouping similar entities together Supervised Segmentation Input: Raw image, Labelled Image Output: Segmented Image Unsupervised Segmentation Input: Raw image Output: Segmented Image 7
- 9. Methodology Methodology consist of following 2 stages: 1. Unsupervised Sementic Segmentation using Clustering algorithms 2. Post-processing using Seed based Region Growing algorithm Input Data ● Satellite: IRS Resourcesat–2 ● Sensor: LISS-IV ● Spatial Resolution: 5.8 m ● No of Bands: 3 ● Bit depth: 16 – 10 bit quantization (1024 different tones can be assigned to a pixel) ● Size: More than 3000 X 3000 ● Cities: Vadodara, Jabalpur, Sagar, Satana, Ujjain, Varanasi, Rampur 9
- 10. Input Images Vadodara Varanasi Ujjain 10
- 11. Raw and Manually Labelled Images 11
- 12. Raw and Manually Labelled Images 12
- 13. Methodology (step-1 of 2) 13
- 14. DBSCAN ● DBSCAN: Density-based spatial clustering of applications with noise ● Its clusters are defined as areas of higher density than the remainder of the data set ● It clusters water bodies accurately in our dataset ● DBSCAN(eps=3.5, min_samples=5, metric=’euclidean’, n_jobs=1) ● eps: The maximum distance between two samples for them to be considered as in the same neighborhood ● min_samples: The number of samples in a neighborhood for a point to be considered as a core point 14
- 15. K-means ● k-means clustering aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster ● KMeans(n_clusters=3, n_init=10, n_jobs=1) ● n_init: Number of time the k-means algorithm will be run with different centroid seeds 15
- 17. Mean Shift 17
- 20. Intuitive Description new color mean value is calculated 20
- 21. Intuitive Description move that window to the area of maximum pixel density 21
- 22. Intuitive Description new center for the next iteration 22
- 23. Intuitive Description iterate until the spatial and the color mean stops changing 23
- 24. Mean Shift ● Mean Shift: Finding modes in a set of data samples, manifesting an underlying probability density function (PDF) in RN ● It is a procedure for locating the maxima of a density function given discrete data sampled from that function ● Thus, it is using a non-parametric density gradient estimation 24
- 27. Non-Parametric Density Estimation ● for each channel, separate colour histograms will be generated 27
- 28. Mean Shift Results Down Sampled by: 0 1 2 3 28
- 29. N-Cut (Graph-Cut) ● Given an image or image sequence, set up a undirected weighted graph G = (V; E) ● The nodes are pixels ● The weight on the edge connecting two nodes is the measure of the similarity between the two nodes in terms of colour, texture etc. ● The objective of normalized partitioning is to optimize the cut value 29
- 30. Watershed ● The watershed transformation treats the image it operates upon like a topographic map ● Intuitively, a drop of water falling on a topographic relief flows towards the "nearest" minimum ● The "nearest" minimum is that minimum which lies at the end of the path of steepest descent 30
- 31. Discussion ● DBSCAN: clustered only water bodies ● K-mean: loss of data ● Mean shift: overall good result ● N-cut: poorest result (time complexity is too high) ● Watershed: not good for this dataset Reference [1] 31
- 32. Methodology (step-2 of 2) 32
- 34. Region Growing ● Group pixels or sub-regions into larger regions when homogeneity criterion is satisfied (assumed criterion is a range of pixel value in all 3 bands) ● Region grows around the seed point based on similar properties (grey level, texture, color) (9 seed points have been assumed) ● It is better in noisy image where edges are hard to identify 34
- 36. Output of RG on Mean Shift & Raw Image As of now, step 2 has directly been performed on raw image. But after tweaking the parameters, output of mean shift can be used 36
- 37. Results Manually Labelled Input Output Of RG Vadodara City Jabalpur CitySagar City 37
- 38. Results Manually Labelled Input Output Sagar City Jabalpur City 38
- 39. Results Manually Labelled Input Output Jabalpur City Jabalpur City 39
- 40. Conclusion ● As of now, RGB Threshold of 14 and B&W pixel voting of 50% applied on RAW image without preprocessing gives best result (based on subjective assessment) ● I will now tune parameters for unsupervised segmentation and try to arrive at results which are better than using RAW image 40
- 41. Future Work ● Twiking of Hyperparameters such as RGB Threshold value and Voting value of B&W pixels for better accuracy ● Implementation of Progressive Thresholding technique to improve voting Labelled Input Output Region Growing Voting Value (Fix 50%) RGB Threshold (10,15,20) Feedback output image as new labelled image 41
- 42. References [1] Xia, Xide, and Brian Kulis. "W-Net: A Deep Model for Fully Unsupervised Image Segmentation." arXiv preprint arXiv:1711.08506 (2017). [2] Shi, Jianbo, and Jitendra Malik. "Normalized cuts and image segmentation." IEEE Transactions on pattern analysis and machine intelligence 22.8 (2000): 888-905. [3] Zhou, Yong-mei, Sheng-yi Jiang, and Mei-lin Yin. "A region-based image segmentation method with mean-shift clustering algorithm." Fuzzy Systems and Knowledge Discovery, 2008. FSKD'08. Fifth International Conference on. Vol. 2. IEEE, 2008. 42
- 43. Thank you!