Improvement of salient-region detection using an integrated bottom-up model

Modeling visual attention is a challenging task for machine vision. In this paper, inspired by the mechanism of human visual system, we propose an integrated model to detect generic salient-regions in a purely bottom-up manner. Instead of only employing early visual features in most relevant works, the saliency of discriminative local regions is also conducted to represent the spatial entropy, which is believed as a significant aspect of the selective attention. The final visual saliency can be detected by combining these two complementary and independent mechanisms. To demonstrate the effectiveness and robustness, both qualitative and quantitative experiments are designed. The results show that the proposed model can achieve satisfying performances, even in highly cluttered scenes.