Presentation is loading. Please wait.

Presentation is loading. Please wait.

Image Segmentation Region growing & Contour following Hyeun-gu Choi Advisor: Dr. Harvey Rhody Center for Imaging Science.

Similar presentations


Presentation on theme: "Image Segmentation Region growing & Contour following Hyeun-gu Choi Advisor: Dr. Harvey Rhody Center for Imaging Science."— Presentation transcript:

1 Image Segmentation Region growing & Contour following Hyeun-gu Choi Advisor: Dr. Harvey Rhody Center for Imaging Science

2 5.8.98Hyeun-gu Choi Background Definition of image segmentation –Subdivides an image into its constituent parts or objects. Region growing –Based on similarity of gray-level values Contour following –Based on discontinuity of gray-level values

3 5.8.98Hyeun-gu Choi Detection of discontinuity Gradient operators - First derivative(Sobel) –Zero value - constant gray level –Positive value - transition from dark region to light region –Negative value - transition from light region to dark region

4 5.8.98Hyeun-gu Choi Detection of discontinuity Gradient operators - Second derivative (Laplacian) –Positive value - dark region –Negative value - light region –This operator is good for deciding whether a pixel is on the dark or light side of an edge.

5 5.8.98Hyeun-gu Choi Threshold Selection Based on Boundary Characteristics Application of these operators (1) T: Threshold value s(x,y): result image

6 5.8.98Hyeun-gu Choi Examples of derivative operator Original image with background picture Segmented image without background by Sobel and Laplacian operator

7 5.8.98Hyeun-gu Choi Results of derivative operator Selected threshold values are around 70. Too small values of threshold mean small change of gray level and too big threshold value may gives noise. Original imageSegmented image Original imageSegmented image

8 5.8.98Hyeun-gu Choi Segmentation by thresholding Simple Global Thresholding –Simply select a threshold value between peak values from a histogram plot and set zero below than threshold and set 255 greater than threshold. –Disadvantage: lost a lot of data information and there is no guarantee of grouped (well- separated) histogram.

9 5.8.98Hyeun-gu Choi Segmentation by thresholding Multiple Thresholding –Select useful region by selecting two thresholds –Simple and very good results –Disadvantage - No guarantee for grouped histogram. Not good for automated system because user should decide threshold value

10 5.8.98Hyeun-gu Choi Examples of multiple thresholding (2) Original spin density image Enhanced image by histogram equalization Segmented image (CSF tissue) 70 - 230 * CSF (Cerebrospinal fluid)

11 5.8.98Hyeun-gu Choi Examples of multiple thresholding Th = 83~255 Th = 0~82 Noise

12 5.8.98Hyeun-gu Choi Label Region algorithm Basic Concept –Scan through an image pixel by pixel –Compare the gray value of center pixel with those of top and left pixels. –If compared gray values are the same, those pixels are categorized to one group. –After scanning all pixels, pixels with the same gray level value will be grouped. Disadvantage - Very vulnerable to noise.

13 5.8.98Hyeun-gu Choi Label Region algorithm Result of label region algorithm Modified label region algorithm Modified label region algorithm has some margin. That is, when pixel values are compared, only if pixel value difference is bigger than some value, the algorithm classify the pixel to new class. Modified algorithm is less vulnerable to noise. Original image

14 5.8.98Hyeun-gu Choi Hough Transform (1) Consider a point (x i, y i ). General equation for this point is y i = ax i + b or b = -x i a+y i If two points (x i, y i ) and (x j, y j ) are on the same line two points have an intersection in parameter space (ab plane). x i,y i x j,y j x y a b b = -x i a+y i b = -x j a+y j a` b` xy plane Parameter space

15 5.8.98Hyeun-gu Choi Problem - Both slope and intercept approach infinity as the line approaches the vertical. Solution - change to  plane y i = ax i + b  x cos  + y sin  =  Hough Transform Normal representation of a line x y  

16 5.8.98Hyeun-gu Choi Subdivision of the parameter space into so- called accumulator cells and counting intersection points. Intersection point with a  value and a  value means two data points are on the same line with  and . Hough Transform  max  max    min  min Quantization of the  plane into cells. (x 1, y 1 ) (x 2, y 2 )

17 5.8.98Hyeun-gu Choi Hough Transform Four dots in xy plane Result of Hough Transform Four dots in parameter space Intersections found by accumulation cell 1 1 2 2 3 3

18 5.8.98Hyeun-gu Choi Disadvantage of Hough Transform –Vulnerable to noise –Difficult to find lines in complicated images –Difficult to change the shape which looking for Line : y = ax + b Circle : (x-C 1 ) 2 + (y-C 2 ) 2 = C 3  3Dim parameter space –Long processing time About 25 minutes for 192 x 128 size image by Pentium 166MHz, 48Mbyte RAM Hough Transform

19 5.8.98Hyeun-gu Choi Hough Transform Original imageGradient image Found diagonal lines because diagonal direction has a lot of data points Gradient image in parameter space Found intersections

20 5.8.98Hyeun-gu Choi Hough Transform Original image Thresholded image Gradient image threshold Histogram of original image

21 5.8.98Hyeun-gu Choi Gradient image in parameter space Found intersections Found linesThresholded image Hough Transform

22 5.8.98Hyeun-gu Choi Graphic User Interface (Widget) A graphic user interface is created for demonstration of segmentation methods in IDL (Interactive Data Language). Widget interface is consisted of five image windows and one plot window. 11 full down main menus. Advantage - Easy to organize many algorithms and easy to modify.

23 5.8.98Hyeun-gu Choi Graphic User Interface (Widget)

24 5.8.98Hyeun-gu Choi Main menus and sub menus –Files New - Get an new image. Save Image –Switch - User can switch two windows –Filter High pass, low pass, unsharp, laplacian, vertical edges, horizontal edges, sobel, median, and custom. –Arithmetic - Add and subtract two windows. Graphic User Interface (Widget)

25 5.8.98Hyeun-gu Choi –Add Noise –Label Region Label region (IDL library) and Modified label region –Histogram Histogram of source window, Histogram equalization, one threshold, and two thresholds –Color table - Emphasize the processed image. Graphic User Interface (Widget)

26 5.8.98Hyeun-gu Choi –Hough Manual - User can choose intersections by clicking in parameter space. Auto - Computer do the whole process. –Zoom - Zoom in an image with several magnification (2x, 3x, 4x). –Done - Finish the interface. Graphic User Interface (Widget)

27 5.8.98Hyeun-gu Choi Reference (1) Gonzalez, Rafael C. and Woods, Richard E.[1992] “Digital Image Processing” Addison- Wesley Publishing Company (2) Lynn M. Fletcher, John B. Barsotti, Joseph P. Hornak [1993] “A Multispectral Analysis of Brain Tissues” Magnetic Resonance in Medicine 29:623-630


Download ppt "Image Segmentation Region growing & Contour following Hyeun-gu Choi Advisor: Dr. Harvey Rhody Center for Imaging Science."

Similar presentations


Ads by Google