Markov chain based computational visual attention model that learns from eye tracking data

• We use Markov chain to model the visual attention.
• Our visual attention model is based on low level and high level image features.
• We use the real eye tracking data to train our visual attention model.
• We measure performances of attention models by comparing them with human fixations.
• Our model is more consistency with the attentional deployment of humans.

Computational visual attention models are a topic of increasing importance in computer understanding of images. Most existing attention models are based on bottom-up computation that often does not match actual human attention. To address this problem, we propose a novel visual attention model that is learned from actual eye tracking data. We use a Markov chain to model the relationship between the image feature and the saliency, then train a support vector regression (SVR) from true eye tracking data to predict the transition probabilities of the Markov chain. Finally, a saliency map predicting user’s attention is obtained from the stationary distribution of this chain. Our experimental evaluations on several benchmark datasets demonstrate that the results of the proposed approach are comparable with or outperform the state-of-art models on prediction of human eye fixations and interest region detection.