M.Sc Thesis

M.Sc StudentGolubchyck Roman
SubjectImproving the Saliency Algorithm by Grouping Cues
DepartmentDepartment of Computer Science
Supervisor PROF. Michael Lindenbaum


The human vision system can discriminate between the important features in an image (denoted figure) and unnecessary features (denoted background). It also can divide the figure features into groups such that each group represents some image structure, corresponding to some object in the scene.

The goal of computerized perceptual organization is to acquire the same ability. The saliency algorithm, the starting point of this research, aims to find long smooth curves passing through image edges, and is indeed able to extract important boundaries and divide them into subsets. The algorithm's results depend on the grouping cue, which may be interpreted as the probability of connecting two feature points in the curve. According to this interpretation, the algorithm finds the curves associated with maximal expected length. To this end, it calculates saliency values at every edge. The main goal of this thesis is to achieve reliable grouping cues, for which the saliency values of figure features are much larger than those of background features. We start from simple learning techniques that are based on local data for computing grouping cues. Then, we use saliency performance analysis for computing grouping cues, which should optimally separate between saliency distributions of figure and background features. Afterwards, we use both local and global data for achieving more reliable cues. Finally, we propose the grouping cues for propagating curves via occlusions in the image. Our results show that:

1. Simple learning techniques for grouping cue computation can significantly improve the performance of the saliency algorithm.

2. The performance analysis of the generalized saliency algorithm can be used for building the optimal grouping cue - the cue that best separates figure and background saliency distribution.

3. A grouping cue, which is based on both local and global information, is more useful than that based on local information only.

4. Merging the saliency process over different scales helps in solving the problem of occlusions.