Multi-Scale Amplitude Spectrum Substitution for Visual Saliency Detection
Fangbin Xu, Dongyue Chen, Xingya Chang · 2021
Saliency detection is among the fundamental issues in simulating human visual mechanism. In this paper, we propose a new framework called Amplitude Spectrum Substitution (ASS) to summarize typical frequency-based saliency models. We then present a saliency detection model by applying the ASS method to a multi-scale structure, where substituted amplitude spectrums of all scales are trained from plenty of natural images. We further provide a novel measurement of global saliency called Salient Feature Number (SFN) to find out the optimal saliency map in the proper scale. Experimental results clearly demonstrate that the proposed model outperforms fifteen state-of-the-art competing models in predicting human fixations.