Chinese Journal of Computers   Full Text
Title	A Perceptual Grouping Algorithm Based on Global Salient Structure
Authors	ZOU Qi LUO Si-Wei ZHONG Jing-Jing
Address	(School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044)
Year	2007
Issue	No.11(2008—2016)
Abstract &   Background	Abstract A grouping algorithm based on global salient structure is proposed. Grouping cues are topological properties namely parallelism and closure and local principles namely proximity and continuity. The most salient edge according to probability reference is selected as grouping seed. Edges determined by global statistical dependency are selected as subsequential ones with the most probability of being in the same group with the seed. In perceptual grouping process, attention is employed in grouping to both reduce optimal space and decide pop-out sequence of groups according to their salience. Compared with algorithms adopting local salient relations, above algorithm provides more reliable cues for nature images. This group-based attention makes the effect close to human perception. Experiments on Berkley image database show above algorithm achieves accuracy competitive to Ncut and mini-cut algorithms. It reaches lower error rate and missing rate especially on images with litter texture. Meanwhile, compared with graph cut methods grouping on pixels, the proposed algorithm grouping on edges reduces input dimensionality, therefore less restrictive in image size. keywords perceptual organization; topological property; attention background The subject is supported by two projects. One is project of National Natural Funds (grant No.60373029), Effective Coding Model based on Human Vision Perception. The other is project of Department of Education Funds for Doctor (grant No.20050004001), Research on Mechanism and Algorithms of Perceptual Organization. The former project adopts the method combining cognitive science and neural science. The target is to advance new cognitive model fitting theoretical framework of effective coding in human vision cognitive system better. Through analyzing cooperation mechanism among multi-vision modules including attention and memory, the authors construct a hierarchical and parallel model, which can approach information processing like human brain to some extent. The results are to be validated by being applied into practical system such as segmentation, grouping and recognition. The latter project researches cognitive mechanism and corresponding computational models of perceptual organization and further applies the models into intelligent information processing system such as automatic object detection and recognition. The approach presented in this paper focuses on middle stage of visual information processing-perceptual grouping. Statistical method is employed to organize primitive features from low level into objects, which reflect global structure with scenery meaning. Thus it provides objects for advanced processing in high level. Computation and representation of topological relations are key to grouping algorithms. Existing methods express topological relations through local features or linear combinations of them, which result in bad effect. This paper defines a global topological relation-closure by rigorous probability reasoning and parallelism. So it provides more reliable cues for grouping. Besides, attention is employed to both enhance the model’s plausibility in biology way-decided pop-out sequence of groups according to their salience, and improve computational feasibility-reduce complexity.