A unified method for comparison of algorithms of saliency extraction

Extracting salient regions of a still image, which are pertinent areas likely to attract subjects' fixations, can be useful to adapt compression loss according to human attention. In the literature, various algorithms have been proposed for saliency extraction, ranging from region-of-interest (ROI) or point-of-interest (POI) algorithms to saliency models, which also extract ROIs. Implementing such an algorithm within image sensors implies to evaluate its complexity and performance of fixation prediction. However, there have been no pertinent criteria to compare these algorithms in predicting human fixations due to the different nature between ROIs and POIs. In this paper, we propose a novel criterion which is able to compare the prediction performance of ROI and POI algorithms. Aiming at the electronic implementation of such an algorithm, the proposed criterion is based on blocks, which is consistent with processing within image sensors. It also takes into account salient surface, an important factor in electronic implementation, to reflect more accurately the prediction performance of algorithms. The criterion is then used for comparison in a benchmark of several saliency models and ROI/POI algorithms. The results show that a saliency model, which has higher computational complexity, gives better performance than other ROI/POI algorithms.