sandeepkonam_statisticalanalysisofimageprocessingtechniques for object counting
TRANSCRIPT
Motivation
Counting !
Source (Clockwise) :Public domain image (created by the Dartmouth Electron Microscope Facility) ; retrieved from www.bloodwork.com ; Photo by Marty Snyderman
© 2014 Sandeep Konam . All rights reserved.
Related Work
• Hough Transform
• RANSAC
• Neural Network
• Shape Matching
© 2014 Sandeep Konam . All rights reserved.
Related Work
• Hough Transform
• RANSAC
• Neural Network
• Shape Matching
© 2014 Sandeep Konam . All rights reserved.
Related Work
Hough Transform
40 out of 50 circles detected as indeed circles© 2014 Sandeep Konam . All rights reserved.
Related Work
• Hough Transform
• RANSAC
• Neural Network
• Shape Matching
How? Randomly chooses pairs of edges to form a line hypothesis and then testhow many other edges fall onto this line
Fallout : Many more hypotheses may need to be generated and tested than those obtained by finding peaks in the accumulator array.
© 2014 Sandeep Konam . All rights reserved.
Related Work
• Hough Transform
• RANSAC
• Neural Network
• Shape Matching
How? Through pre-setting of similar shape templates and training them.
Fallout : System needs to be trained with all the variants and discrepanciesassociated with the shapes.
© 2014 Sandeep Konam . All rights reserved.
Related Work
• Hough Transform
• RANSAC
• Neural Network
• Shape Matching
How? Requires the shape description and representation based on which the image containing shapes to be detected and the template are compared.
Fallout : Difficult to get a fit descriptor.
© 2014 Sandeep Konam . All rights reserved.
Algorithm
Random Image Generation
Introducing random quotient
Coordinate Selection
Drawing shapes based on Mathematical
Properties
Avoid shapes around the corners – Else Distinctness
of the shapes might be lost
Vertices and angles need to be fed for closed
shapes other than line, circle and ellipse
© 2014 Sandeep Konam . All rights reserved.
Algorithm
Detection and Probability
Find contours
Size of contours
Shape Classification
Image is split into parallel regions
Probability in the vicinity is calculated
Depending on the size, classification of shapes can be
done
Angles and other metrics might be necessary to detect and
classify complex shapes
© 2014 Sandeep Konam . All rights reserved.
Buffon’s Needle Problem
p(C) = 2lπd
“ Let a needle of length l be thrown at random onto a horizontal plane ruled with parallel straight lines spaced by a distace d from each other, with d > l. What is the probability p that the needle will intersect one of these lines? “
Generalized Buffon’s Needle Problem : p(C) = aπd
• Limiting argument to Circle : p(C) = 2π𝑟πd
= 2𝑟d
© 2014 Sandeep Konam . All rights reserved.
Results
Needles
Exp. No. N n p(S) p(C) % Abs. Error1 100 38 0.380 0.455 7.5
2 150 61 0.406 0.455 4.83
3 200 84 0.420 0.455 3.5
4 400 172 0.43 0.455 2.5
5 600 264 0.44 0.455 1.5
6 800 355 0.444 0.455 1.13
7 850 380 0.447 0.455 0.794
© 2014 Sandeep Konam . All rights reserved.
Results
Circles
Exp. No. N n p(S) p(C) % Abs. Error1 50 41 0.82 0.857 3.7
2 100 84 0.84 0.857 1.7
3 150 127 0.846 0.857 1.1
4 200 170 0.85 0.857 0.7
5 250 213 0.852 0.857 0.5
© 2014 Sandeep Konam . All rights reserved.
Conclusion
• As the number of mathematical shapes increase, results of proposed algorithm converges to that of empirical calculations
Future Scope
• Efficient Algorithm needs to be developed for overlapping images
• Developed algorithms are to be applied on real time images
© 2014 Sandeep Konam . All rights reserved.