Object segmentation by attention-induced oscillations

We propose a network architecture based on spiking neurons performing visual object segmentation. The model is able to deal with graded inputs as gray-level images. In a feedforward mode, neurons in a feature map encode the input by their average firing rate. A saliency map detects high-contrast regions of the input and an attention map creates a feedback signal sequentially enhancing all the salient regions. The feedback signal modulates the firing pattern of feature map neurons by: 1) modifying their short-term spiking frequency, and 2) synchronizing all neurons of the same object, while asynchronizing the others. Bursts of high frequency alternate with the basic frequency determined by the external input