A Jeffrey divergence based irregular pyramid method for pre-attentive visual segmentation

This paper proposes a pre-attentive visual segmentation method by using Jeffrey divergence based irregular pyramid, termed as “JIP” algorithm. The proposed JIP algorithm obtains a suboptimal labeling solution under the condition that the number of segments is not manually given. This algorithm models each node at higher pyramidal levels as a probabilistic distribution, based on which Jeffrey divergence is employed to measure the inter-node distance while entropy is used to measure the intra-node distance. As a result, the neighborhood system for the graph of each higher pyramidal level can be automatically set up according to the similarity and graphic constraints. During the hierarchical accumulation of the irregular pyramid, the segments can emerge once they are represented by single nodes at certain levels. Experimental results have shown that this proposed JIP algorithm outperforms other benchmark segmentation algorithms in terms of segmentation accuracy and labeling cost.