Related Experiment Videos
An efficient parameterless quadrilateral-based image segmentation method.
Ronald H Y Chung1, Nelson H C Yung, Paul Y S Cheung
1Department of Computer Science, The University of Hong Kong, Room 425, Chow Yei Ching Bldg., Pokfulam Road, Hong Kong. hychung@cs.hku.hk
IEEE Transactions on Pattern Analysis and Machine Intelligence
|September 22, 2005
Summary
This study introduces a novel quadrilateral-based image segmentation framework. This method efficiently merges similar regions, offering accurate segmentation without over or undersegmentation for better content-based applications.
Area of Science:
- Computer Vision
- Image Processing
- Computational Geometry
Background:
- Traditional image segmentation methods often struggle with noise and detail, leading to over or undersegmentation.
- Existing techniques like seeded region growing, K-means, and constrained gravitational clustering have limitations in accuracy and efficiency.
Purpose of the Study:
- To propose a general quadrilateral-based framework for robust image segmentation.
- To develop an efficient, parameterless algorithm derived from this framework.
- To demonstrate superior performance compared to existing segmentation techniques.
Main Methods:
- Constructing quadrilaterals from an edge map.
- Merging neighboring quadrilaterals with similar features.
- Developing a parameterless algorithm for efficient region description.
- Evaluating segmentation accuracy and object representation.
Main Results:
- The framework enables elimination of local variations for accurate region description.
- Segmented regions are represented by multiple levels of quadrilaterals.
- The proposed algorithm achieves accurate segmentation without significant over or undersegmentation.
- The algorithm outperforms seeded region growing, K-means, and constrained gravitational clustering.
Conclusions:
- The quadrilateral-based framework provides an effective approach to image segmentation.
- The derived algorithm offers high performance and accurate region representation.
- This method is well-suited for content-based image retrieval and analysis applications.