Automatic location of optic disk in retinal images

1
Automatic location of optic disk in retinal images

• Author Huiqi Li, Opas Chutatape
• Source Proceeding of IEEE ICIP 2001
• Speaker Shi-Fang Yang-Mao

2
Outline

• Candidate region determination.
• PCA method.
• (Eigenfaces for recognition, 1991)
• Results and discussion.
• Conclusion

3
Candidate region determination

• The pixels with the highest 1 gray level in
  intensity image are selected and clustered.
• Some cluster take as noises if has less than 100
  pixels.
• Dimensions of optic disk is about 65100 pixels.

4
PCA method (1/5)
eigen-space
Training set of optic disk images
1. Training
3. Recognize
2. input image
5
PCA method (2/5)

• Training stage
• Pick the training set of optic disk images.
• Compute the average and each difference between
  training set images and the average.
• Build the eigen-space.
• Recognize stage
• Project the sub-image of input image into
  eigen-space.
• Calculate the Euclidian distance of the input
  image and the projection.
• Center of the image with minimum Euclidian
  distance is the location of optic disk.

6
PCA method (3/5)

• Pick the training set of optic disk images.
• Compute the average and each difference between
  training set images and the average.

training set images
each difference
average
7
PCA method (4/5)

• Calculate the eigen-space.

find the eigenvector ?
and
Let
eigenvector
eigenvalue
eigenvector
Assume
eigenvalue
, M lt M
Eigenvector is the Linear combination of each
difference between training set images and the
average
forms the eigen-space
8
PCA method (5/5)

• Find the minimum Euclidian distance from the
  input image( is only a subimage from input
  image and ).

Compute the weight of eigenvector
Project image onto eigenspace
Calculate Euclidian distance
Center of the sub-image with minimum Euclidian
distance is the location of optic disk.
9
Results and discussion
(1) Candidate region
(3) Locate the optic disk
(2) Euclidian distance
(4) Better method
10
Conclusion

• Face recognize method PCA used here.
• Statistics method used is possible.
• Classify the training set image get better result
• LDA/FDA maximize the between-class scatter and
  minimize the within-class scatter.
• SVM