1 / 23

A Model of saliency-Based Visual Attention for Rapid Scene Analysis

A Model of saliency-Based Visual Attention for Rapid Scene Analysis. Reporter: You Jian. Laurent Itti, Christof Koch, and Ernst Niebur. Introduction. Visual attention: Focus of attention Focus selection Rapid, saliency-driven, task-independent Slow, volition-control, task-dependent. Model.

mortimer
Download Presentation

A Model of saliency-Based Visual Attention for Rapid Scene Analysis

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Model of saliency-Based Visual Attention for Rapid Scene Analysis Reporter: You Jian Laurent Itti, Christof Koch, and Ernst Niebur

  2. Introduction • Visual attention: • Focus of attention • Focus selection • Rapid, saliency-driven, task-independent • Slow, volition-control, task-dependent

  3. Model

  4. Model • Size: usually 640 * 480 • 9 [0..8] spatial scales are created. • Dyadic Gaussian pyramids • Center is a pixel at scale • Surround is the pixel at scale • :interpolation to the finer scale and point-by-point subtraction.

  5. Extraction of Early Visual Features • r, g, and b is the red, green, and blue channels of the input image. • Intensity image: • the r, g, and b channels are normalized by I in order to decouple hue from intensity. • Hue variation are not perceivable at very low luminance • Normalization is only applied at the location where Other location yield zero r, g, b

  6. Four broadly-tuned color channels

  7. Create Gaussian Pyramids • Create Gaussian Pyramids for I, R, G, B, and Y • and

  8. Feature Maps-Intensity Contrast • Feature Map • Center-surround differences( ) between a “Center” fine scale c and a “surround” coarser scale s yield the feature maps. • A set of six maps

  9. Feature Maps-Color Double Opponent • Spatial and chromatic opponency in human primary visual cortex. • red/green and green/red • blue/yellow, and yellow/blue color pairs

  10. Feature Maps-Orientation • Obtained from I using oriented Gabor pyramid • In total 42 feature maps are computed : • six for intensity • 12 for color • 24 for orientation.

  11. The Saliency Map • The saliency map • Represent the conspicuity-or “saliency”-at every location in the visual field by a scalar quantity • Guide the selection of attended locations, based on the spatial distribution of saliency.

  12. The Saliency Map • Combination of the feature maps • Provides vides bottom-up input to the saliency map • Modeled as a dynamical neural network. • Difficulty: • Different dynamic ranges • Different extraction mechanisms. • Salient objects appearing strongly in only a few maps

  13. The Map Normalization • Normalization: • Promotes maps with a small number of strong peaks of activity (conspicuous location). • Suppress maps with numerous comparable peak response.

  14. The Map Normalization • The consist of:

  15. The Map Normalization • Biological motivation • Lateral inhibition mechanisms. • Neighboring similar features inhibit each other via specific, anatomically defined connections.

  16. Combination of Feature Maps • Feature maps are combined into three ”conspicuity maps”: • for intensity • for color • for orientation

  17. Combination of Feature Maps • Combine operator consist of : • reduction of each map to scale four • point by point addition

  18. Combination of Feature Maps • The motivation of the creation of the three separate channels and the individual normalization: • Similar features compete strongly for saliency, • Different modalities contribute independently to the saliency map.

  19. Focus Selection • Model the SM as a 2D layer of leaky integrate-and-fire neurons at scale four • The potential of SM neurons at more salient location hence increases faster • Each SM neurons excites its own WTA neuron. • WTA neurons independent until one first reach the threshold and fires.

  20. Focus Selection • That triggers three simultaneous mechanism: • FOA is shifted • Global inhibition of the WTA • Local inhibition of the Winner

More Related