Image segmentation based on edge and corner detectors

1
Image segmentation based on edge and corner
detectors

• Joachim Stahl
• 04/21/2005

2
Problem

• The results of edge based image segmentation are
  affected by the performance of the underlying
  edge detector.
• In particular, edge detectors are weak at corner
  points.
• Solution Introduce the results of a corner
  detector to make up for this weakness.

3
Image Segmentation

• Goal to separate an image into foreground and
  background
• Foreground represents an object of interest.
• Different approaches, pixel-based and edge-based.

4
Edge Detection

• Returns a binary image indicating pixels in the
  original image where an abrupt changes in pixel
  intensity occur.
• Most famous method, Canny edge detector.
• One major drawback it works poorly at corners.

5
Corner Detection

• Like edge detection, deals with points of high
  intensity changes, but also with abrupt changes
  in the direction of the edge track.
• Most famous method, Harris corner detector.

6
Basic Idea

• In order to create connected boundaries, edge
  grouping methods create fragments to fill the
  gaps.
• Let the construction of these gap-filling
  fragments be affected by the presence of a corner
  point.

7
Expected improvement

• By having this extra corner information
  incorporated, it is expected that a method can
  improve its detection in these special cases

8
Considering Track Length

• Another important consideration, the length of
  the edge track to which the fragment belongs.
• Not considering this length could lead to false
  corners due to inaccurate edge approximation
  introduced by noise in the image.

9
Sample application

• Implement method around Ratio Contour.
• RC assigns a cost to each fragment based on
  proximity and continuity.
• Let the presence of a corner affect the curvature
  of nearby fragments. We can reformulate the cost
  function of RC to reflect this

10
Result
11
Conclusions

• Incorporating corner detection to edge grouping
  can improve results.
• However, improvement is only noticeable in a
  small percentage of cases.
• Still a work in progress.
• It is a motivation to also work on edge detection
  improvement.

12
Thank you!

• Questions?