Saliency detection based on Markov chain and Adaboost

Abstract

In this paper, an effective saliency detection method is proposed to minimize the error caused by boundary prior. When the salient objects appear in picture edge, the results generated by the traditional methods are usually not ideal. We proposed a two-step algorithm to optimize the saliency detection. Firstly, an initial saliency map is generated by absorbing Markov chain model. In this model, super-pixels in edge of the picture are treated as virtual boundary absorbing nodes. The remaining super-pixels are treated as transient nodes. Transient nodes' absorbed time are calculated as their saliency value. Then the initial saliency map is utilized as training samples to train an Adaboost classifier. The initial saliency map is refined with the trained Adaboost model to get a optimized saliency map. By combining the two maps we get the final saliency map. Experiments show that our algorithm is superior to other traditional saliency detection method when salient objects are close to image boundaries.