INTERACTIVE IMAGE SEGMENTATION BASED ON ACCELERATED LEARNING USING PAIRWISE CONSTRAINTS

Segmentation often is used as a pre-processing step before higher level image processing or understanding. Manually segmenting images can be laborious and time-consuming, especially for large images, and can be prone to subjectivity. At the same time, developing a generic unsupervised segmentation algorithm that can be accurately applied to all, or even many, types of images is not straight-forward. This is mainly because automatic image segmentation is known to be an ill-posed problem in the sense that images can be segmented differently depending on hard-to-specify high level goals. Our work in this paper is based on constrained spectral clustering that iteratively incorporates user feedback by propagating it through the calculated affinities. The original framework does not scale well to large data sets, and hence is not straightforward to apply to interactive image segmentation. In order to address this issue, we adopt advanced numerical methods for eigen-decomposition implemented over a sub sampling scheme. Our key innovation, however, is an active learning strategy that chooses pair wise queries to present to the user in order to increase the rate of learning from the feedback. Performance evaluation is carried out on the Berkeley segmentation and Graz-02 image data sets, confirming that convergence to high accuracy levels is realizable in relatively few iterations.