Selective Visual Attention: Computational Models and Applications

3.4 Graph-based Visual Saliency

An alternative approach is based on a parallel graph calculation model called graph-based visual saliency (GBVS) [6]. The first two steps of the GBVS model are similar to the original BS model that extracts low-level features from input image by various filters, but there are differences such as: (1) the centre–surround process in the alternation is omitted for orientation feature channels, as mentioned in Step 2 of Section 3.3.1; (2) for the contrast feature map of intensity channel, the luminance variance (centre–surround) is computed in a local neighbourhood of the same scale; (3) the number of scale is changed (for simplicity, only a few scales are considered); (4) Multiscale feature maps are rescaled to the same size with different resolutions. Notably, in the GBVS model, creating the feature maps does not need subtraction between the finer and coarser scales as in the BS (Equation 3.9a). The next steps of the GBVS model are to form the activation map from each feature map based on graph theory (i.e., by using the adjacency matrix of the graph and iterations of the Markov matrix), and normalizing of the activation map is also based on the Markov matrix, and the two steps are totally different from the original BS model.

3.4.1 Computation of the Activation Map

Suppose a feature map p (in the form of greyscale image) is given, , an activation map ...