- Title:
- Visual Perception Model and Object Segmentation
- Presenter:
- Professor Philip Ogunbona
- Date:
- February 18, 2005
- Abstract:
- The extraction of semantically meaningful objects from single image or a sequence of images has recently become an active research topic in multimedia signal processing. Algorithms developed so far for object segmentation, in general, fall into three categories, each category in fact adopts different definition of ”object”. The first one, known as moving object segmentation from a sequence of images defines the object as a group of pixels moving in the same or coherent manner. The second category is an extension of traditional region or edge based segmentation with some heuristic about the formation of the objects or with some extra information such as depth or range. The region or edge based segmentation usually serves as low level processing and a rule based system is applied to group the segmented regions into objects. Objects are often defined in this case by their geometric or chromatic formation. The third category employs visual perception theory established by gestalt psychologists and cognitive scientists, especially, the principles of perceptual grouping. Research focus in this category has been in the past on how to quantify the qualitative rules of perceptual grouping and how to incorporate these rules into the traditional segmentation scheme. Modeling human visual process is crucial for automatic object segmentation that is able to produce results consistent with human perception.
Based on the latest understanding of how human performs the task of extracting objects from images, we proposed a graph-based computational framework to model the visual process. The model supports the hierarchical nature of human visual perception and consists of the key steps of human visual perception including pre-attentive (pre-constancy) grouping, figure-and-ground organization, and attentive (post-constancy) grouping. A divide-and-conquer implementation of the model based on the concept of shortest spanning tree (SST) has demonstrated the potential of the model for object segmentation.
Enquiries: Wanqing Li (Tel: 4221 5410) or Susan Branch (Tel: 4221 4928).
Back to:
|
