Parsing Images into Region and Curve Processes

Input Image

 

 

 

 

Region Layer

 

Curve Layer

 

Motivation

  1. Curve is a typical pattern appearing in natural images and has been studied extensively in the literature. Also, curves often fall in groups corresponding to high level structures such as railings, trees, and textures, which are key components to be recognized in understanding scenes for visual systems. We study curves as the second types of patterns of interest in the image parsing framework.
  2. Without analyzing curves explicitly, image segmentation algorithms assuming 2D region patterns often have problems in dealing with images with rich 1D structures, for example, parallel railings, weeds, tree branches, etc. Following are some examples by the DDMCMC image segmentation algorithm.

Segmentation results with region patterns only.

These examples show the rationale of introducing curve patterns explicitly.

Results with region and curve patterns .

 

Results

 

Notes:

In this approach, curve pattern is only concerned with degenerated 2D regions, ridge. In other context like SNAKE, curves are some times referred as boundaries of neighboring regions. This type of curves, however, are not independent of region representation where region layer layer is explicitly modeled. Integration of ridge and boundary curves in another view can be seen in a primal sketch, proposed recently by Guo, Zhu and Wu.

 

Publications:

Zhuowen Tu, and Song-Chun Zhu, "Parsing Images into Regions, Curves, and Curve Groups", Int'l Journal of Computer Vision, 2006

Zhuowen Tu and Song-Chun Zhu, "Parsing Images into Region and Curve Processes", 7th European Conference on Computer Vision, May, 2002.