Multi-scale discriminant saliency with wavelet-based Hidden Markov Tree modelling

•Bottom-up saliency can be considered as a binary classification problem between centre and surround classes.•Discriminant power for classification is mutual information between image features and corresponding classes distributions.•A multi-scale structure is integrated into the framework by employing discrete wavelet features and Hidden Markov Tree (HMT).•Standard quantitative tools such as NSS, LCC, AUC and qualitative assessments are used for evaluating the proposed method.•It takes less than 0.4s to process each frame provided that MDIS is run on Intel Xenon Duo-core 2.53 GHz workstation. Supposed saliency is a binary classification between centre and surround classes, saliency value is measured as their discriminant power. As the features are defined by sizes of chosen windows, a saliency value at each location is varied accordingly. This paper proposes computing saliency as discriminant power in multiple dyadic scales of Wavelet Hidden Markov Tree (HMT), in which two consecutive dyadic scales provide surrounding and central features, organized in a quad-tree structure. Their discriminant power is estimated as maximum a posterior probability (MAP) by Expectation-Maximization (EM) iterations. Then, a final saliency value is the maximum discriminant power generated among these scales. Standard quantitative tools and qualitative assessments are used for evaluating the proposed multi-scale discriminant saliency (MDIS) against the well-know information based approach AIM on its image collection with eye-tracking data. Simulation results are presented and analysed to verify the validity of MDIS as well as point out its limitation for further research direction.