Visual Attention Shift based on Image Segmentation Using Neurodynamic System

A method of predicting visual attention shift is proposed based on image segmentation using neurodynamic system in this paper. The input image is mapped to a neural oscillator network. Each oscillator corresponding to a pixel is modeled by means of simplified Wilson-Cowan equations, and is coupled with its 8nearest neighbors. Then the image is segmented by classifying the oscillation curves of the excitatory groups of all the oscillators. The classifier is constructed based on features of frequency, offset, phase and amplitude of the curves. The visual attention shift between the regions on the image is predicted according to the saliency strength of each region. Referring to the mechanism of winner-take-all competition, the saliency of a region is the aggregation of the dissimilarities between this region and all the other ones. Experimental results on images show the effectiveness of our method.