Intelligent Scissor

Base on Sigg95 paper .

Source code

UI control

• "Folder" icon to open and operate on an image.
• First "save" icon to save image with contour.
• Second "save" icon to save the mask generated by the selected contour.
• Third "save" icon to save the area of image cut by the selected contour.
• Image can be zoomed in and out and move around.
• Image has two view mode: original and gradient map.


• To start a new path, press "Put A Seed". Left click to create new seeds, "Delete" to undo previous seed.
  Press "Enter" to end the contour, or double click to close the contour.
• One can hide or show contours as wish.
• To select a particular contour in order to delete it or save the mask, press "Select A Contour".
  Selected contour will show a different color. After selection, press "Delete" to delete it.
• Debug Mode:
  Pixel Node: A cost graph with original image pixel colors at the center of each 3by3 window, and black everywhere else.
  Cost Graph: A cost graph with both pixel colors and link costs. Low cost (darker) for links along image edges.
  Path Tree: See the live expansion of minimum path tree.
  Min Path: Show the minimum path between the current seed and the mouse position.

Implementation

• Basic steps:
  i. Use 8 kernels to apply on each pixel of the image to get a cost matrix (x*y*8 size)
  ii. Do a shortest path search using Dijkstra's algorithm. Insertion and extractMin using priority queue.
• Add a Guassian filter before calculating the image cost matrix.
  ==> Resulting in better adherence at strong edges, but sometimes ignores small (but not is not noise!) disturbances.
• Tried both cost functions given on course webpage and the paper. In my case, the paper one performs better.
• Seed snapping: seed is automatically moved to the closest edge. (search radius 5)
  (Sometimes may cause undesired result but overall performs well.)