Saliency detection in composite stimuli. (a) source image; (b) saliency map of the attention model with invariant features coding. When σ
is 1.8%−3% of the image size (the larger one between the image width and height),σ
is three to four times of σ
, α=1.8, saliency detection results are similar to (b); (c) saliency map of the attention model without invariant features coding, and parameters do not substantially influence the result; (d) saliency map of Itti’s attention model; (e) saliency map of a fully connected network; (f) saliency map of a randomly connected network. The perceptual interpretation of human observers is reproduced only by our model with invariant features.