Saliency Based Image Segmentation

Currently, most of the image segmentation approaches based on region homogeneity can’tsegment objects contained inhomogeneous regions from their surrounding backgrounds. Usually,there are two attempts to overcome such limitation. Firstly, saliency detection can be helpful inextracting the salient objects contained inhomogeneous regions. Secondly, proposed interactive imagesegmentation methods incorporating user interactions. However, there are some shortcomings forthese two directions. Current methods of saliency based image segmentation generate regions thathave poorly defined borders and they fail to homogeneous objects. Meanwhile, interactive imagesegmentation methods are tedious, time-consuming and lacking in precision.According to above two points, this thesis firstly proposed a novel image segmentation approachnamely saliency based N-cut (SNcut). By the use of pixel inhomogeneity factor introduced in INPmethod for the classic normalized cut method, the SNcut exploits the statistical properties of intensityinhomogeneity and gives a good result. Seconly, adopting a newly-proposed idea, i.e., maximalsimilarity based region merging, a framework of saliency-seeded region merging for automaticinteractive segmentation was proposed. The pixels that have different cues but from the same objectare often good candidates for prior interactions, and those pixels at the same time are always withhigher salience attracting human attentions. With the aid of saliency detection, some usefulknowledge about the object of interest and background can be automatically found as priorinteraction.Extensive experiments and comparisons are conducted on a wide variety of natural images.Results show that these frameworks can reliably segment many objects out from their surroundingbackgrounds.