Active Contours Technique in Retinal Image

1
Active Contours Technique in Retinal Image
Identification of the Optic Disk Boundary
Soufyane El-Allali Stephen Brown
Department of Computer Science and
Engineering University of South Carolina Dr. Song
Wang CSCE 790 Spring 2003
2
The Problem

• Objective Using active contours to find the
  optic disk boundary.
• Impediments
• Large image size
• Location of optic disk
• Noise
• Initialization

3
Solution Model
4
Image Pre-processing

• Significance Without pre-processing the active
  contours is strongly influenced by noise.
• Phases
• Thresholding
• Windowing
• Morphological techniques
• Dilation
• Erosion
• Reconstruction

5
Thresholding Windowing

• Threshold Optic disk corresponds to the
  brightest region.
• Gradient marker level is set to obtain the
  threshold.
• Optic disk region corresponds to 245-255 of the
  intensity level.
• Windowing cropped image based on the threshold.

6
Dilation

• Definition Dilation causes objects to dilate or
  grow in size by adding pixels to the boundaries
  of object in an image.
• Dilation depends on a structure element.
• Dilation algorithm.

7
Erosion and Reconstruction

• Definitions
• Erosion causes objects to shrink by removing
  pixels on object boundaries.
• Reconstruction takes the maximum pixel value from
  the original image and the dilated/eroded image.

8
Active Contours Revisted

• Definition Active contours (snakes) is an
  edge-based technique that defines curves within
  an image domain that can move under the influence
  of internal and external forces in order to
  achieve convergence along an object.

is the snakes elasticity
is the snakes regidity
Gaussian functions standard deviation
9
Active Contours Continued

• Objective minimizing the energy functional
• Solution must satisfy Euler Lagranges
  Equation
• Bringing the snakes to equalibrium
• Adding a damping term and an inertial term
• Simple solution using the gradient descent
  algorithm

10
Gradient Vector Flow (GVF)

• Traditional snakes has a tendency not to
  converge in the case of concave shapes.
• GVF Proposed by Xu and Prince
• Static external force h (p, q)
• Minimizes the energy function
• Solution solving the Euler system

is a regularization parameter
11
Experiment

• Traditional snakes
• Before After Preprocessing
• Initialization
• Superposition of GVF fields
• Results

12
Initialization

• Incorrect initialization leads to inaccurate
  results.
• Example
• Snake initialized in an empty GVF field.
• Results in snake resting in same area.

13
Before After Pre-processing
14
GVF fields Superposition

• Motive
• Larger Gaussian standard deviation captures the
  object of interests, yet blurring the edge
  boundary.
• Smaller Gaussian standard deviation stores the
  edge boundary, but does not capture the whole
  object of interest.

• Solution
• Superposing GVF fields with different Gaussian
  standard deviations.

15
Superposition Results
16
Demo
17
Final Results
Original
Final
18
Questions