Extraction of Weighted Saliency Maps in Modelling Bottom-Up Auditory Attention

Hearing is an important part of human daily life. Although humans are exposed to various sounds from different sources and the numbers of receptors of the neural system are limited, they can process complex auditory scenes well. One of the reasons for this human ability is the phenomenon of attention. Auditory attention can be divided into two categories: bottom-up attention and top-down attention. In this paper, a model for simulating the bottom-up attention using weighted saliency maps in the auditory system is proposed. The dataset in this research work is obtained by combining different background noises with the sounds in the ESC database as salient regions, at different SNRs. To evaluate the model, the mean-error criterion was used, which is defined as the time difference between the actual salient point and the salient point detected by the model. The weighted combination of the conspicuity maps of the features using the Genetic algorithm makes the proposed model with an average error of 0.92 seconds to perform better than the baseline model having an average error of 1.91 seconds.