The Study of Edge Detection by Neural Networks

1
The Study of Edge Detection by Neural Networks

• Sang Ki Park
• ECE572 Digital Image Processing (Fall 2004)
• Instructor Dr. Seong G. Kong

2
Order of Presentation

• Neural Network 3
• Motivation 4
• Explain about Work Process 5
• Result 7
• Working Demo
• Discussion 17

3
Neural Network

• Neural Network ?
• NN is consist of Synapse Weights, Biases and
  Neurons
• Input-Output mapping (Black Box Mapping)
• NN is a highly non-linear device and has an
  adaptability
• non linearity of activation function, changeable
  weight
• Multi-Layer Perceptron ?
• Perceptron is a simple form of NN is used for
  classification
• MLP has more than one hidden layer
• Function signal propagate forward
• Error signal propagate backward for correcting
  weight and bias

4
Motivation

• One to one matching
• Flexibility
• Multi-Capacity

5
Training Mask

• Training Data Preparation
• Training data sets were prepared by small
  selected data
• NN can give a generalization in the training
  region (interpolation)

x
x
x
x
x
x
x
x
1st Train
2nd Train
3rd Train
4th Train
6
Compared with Traditional Methods
Sobel Mask
Laplacian Mask
7
Original Image 1
8
Result 1
9
Original Image 2
10
Result 2
11
Original Image 3
12
Result 3


13
Original Image 4
14
Result 4 - 1

15
Result 4 - 2
16
Consideration

• More Training Masks

x
x
curve
x
8
x
x
x
x
x
x
bisect
4
junction
4
x
x
x
x
x
x
cross
x
x
x
x
x
17
Discussion

• More Training Masks make more thick edges
• How to slenderize the thick edge
• Adjust Activation Function (Threshold, Steepness)
• Another Possible Edge Detection Method
• Apply The Radial Basis Neural network to
  Histogram Threshold base Edge Detection

18
Reference

• Suchendra M. Bhandarkar and Hui Zhang, Image
  Segmentation Using Evolutionary Computation,
  IEEE Trans. Evolutionary Computation, vol. 3, no.
  1, pp.1-21, April 1999.
• J. Wesley Hines, MATLAB Supplement to Fuzzy and
  Neural Approaches in Engineering, Wiley
  Interscience, 1997.

19
Th-th-thats All, Folks!