Interactions of Visual Attention and Object Recognition: Computational Modeling, Algorithms, and Psychophysics

• Main Author: Walther, Dirk
• Format: Others
• Published: 2006
• Online Access: https://thesis.library.caltech.edu/895/1/00_DirkWalther_PhDthesis.pdf; https://thesis.library.caltech.edu/895/2/01_DirkWalther_Title.pdf; https://thesis.library.caltech.edu/895/3/02_DirkWalther_Acknowledgments.pdf; https://thesis.library.caltech.edu/895/4/03_DirkWalther_Abstract.pdf; https://thesis.library.caltech.edu/895/5/04_DirkWalther_Contents.pdf; https://thesis.library.caltech.edu/895/6/05_DirkWalther_ListOfFigures.pdf; https://thesis.library.caltech.edu/895/7/06_DirkWalther_ListOfTables.pdf; https://thesis.library.caltech.edu/895/8/07_DirkWalther_Chapter1.pdf; https://thesis.library.caltech.edu/895/9/08_DirkWalther_Chapter2.pdf; https://thesis.library.caltech.edu/895/10/09_DirkWalther_Chapter3.pdf; https://thesis.library.caltech.edu/895/11/10_DirkWalther_Chapter4.pdf; https://thesis.library.caltech.edu/895/12/11_DirkWalther_Chapter5.pdf; https://thesis.library.caltech.edu/895/13/12_DirkWalther_Chapter6.pdf; https://thesis.library.caltech.edu/895/14/13_DirkWalther_Chapter7.pdf; https://thesis.library.caltech.edu/895/15/14_DirkWalther_Chapter8.pdf; https://thesis.library.caltech.edu/895/16/15_DirkWalther_AppendixA.pdf; https://thesis.library.caltech.edu/895/17/16_DirkWalther_AppendixB.pdf; https://thesis.library.caltech.edu/895/18/17_DirkWalther_References.pdf; Walther, Dirk (2006) Interactions of Visual Attention and Object Recognition: Computational Modeling, Algorithms, and Psychophysics. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/P5NY-VC91. https://resolver.caltech.edu/CaltechETD:etd-03072006-135433 <https://resolver.caltech.edu/CaltechETD:etd-03072006-135433>

<p>Selective visual attention provides an effective mechanism to serialize perception of complex scenes in both biological and machine vision systems. In extension of previous models of saliency-based visual attention by Koch and Ullman (Human Neurobiology, 4:219-227, 1985) and Itti et al. (IEEE PAMI, 20(11):1254-1259, 1998), we have developed a new model of bottom-up salient region selection, which estimates the approximate extent of attended proto-objects in a biologically realistic manner.</p> <p>Based on our model, we simulate the deployment of spatial attention in a biologically realistic model of object recognition in the cortex and find, in agreement with electrophysiology in macaque monkeys, that modulation of neural activity by as little as 20 % suffices to enable successive detection of multiple objects.</p> <p>We further show successful applications of the selective attention system to machine vision problems. We show that attentional grouping based on bottom-up processes enables successive learning and recognition of multiple objects in cluttered natural scenes. We also demonstrate that pre-selection of potential targets decreases the complexity of multiple target tracking in an application to detection and tracking of low-contrast marine animals in underwater video data.</p> <p>A given task will affect visual perception through top-down attention processes. Frequently, a task implies attention to particular objects or object categories. Finding suitable features can be interpreted as an inversion of object detection. Where object detection entails mapping from a set of sufficiently complex features to an abstract object representation, finding features for top-down attention requires the reverse of this mapping. We demonstrate a computer simulation of this mechanism with the example of top-down attention to faces.</p> <p>Deploying top-down attention to the visual hierarchy comes at a cost in reaction time in fast detection tasks. We use a task switching paradigm to compare task switches that do with those that do not require re-deployment of top-down attention and find a cost of 20-28 ms in reaction time for shifting attention from one stimulus attribute (image content) to another (color of frame).</p>