Figure 2
From: A neural computational model for bottom-up attention with invariant and overcomplete representation
Diagram of the fully connected network. An image is first convolved with all filters in the first layer, the outputs of which are rectified by the sigmoid and the absolute functions. Then, all the rectified outputs are pooled in the second layer to produce a single feature map, which is subjected to the inhibition simulated by convolution with the DoG function. The final output is the saliency map.