Abstract

F. Tehrani, L. Itti, C. Koch, Visual Search Asymmetries Reproduced by Simple Model, In: Investigative Ophthalmology and Visual Science (Proc. ARVO 2000), Vol. 41, No. 4, p. S423, Mar 2000. (Cited by 1)

Abstract: Purpose: In some instances of visual search, where human observers detect the presence or absence of a special ``target'' visual pattern in an array of identical ``distractor'' patterns, search time asymmetries have been reported when target and distractor patterns are interchanged. A classical explanation for this finding is that targets which contain ``richer features'' are easier to find among simpler distractors than the opposite. For example, a curved line segment is detected faster among straight segments than the opposite, presumably because it has the added property of curvature. Here we test whether a simple computational model of bottom-up attention reproduces such asymmetries. Methods: Our model implements a number of simple multiscale ``feature maps'' selective for colors, orientations and intensity, and combines them into a unique topographic ``saliency map'' which guides attention (Itti et al., IEEE-PAMI, 1998). Stimulus arrays were generated by an automatic program. Search elements were randomly jittered by up to 60% of their size and rotated by up to +/-10deg. Uniform color speckle noise with 10% probability was finally added. For twenty target/distractor pairs (e.g., ``Q'' among ``O'', or open among closed cicles), we generated twenty instances of arrays containing 4x4 to 10x10 elements (seven sizes in total). The resulting 5600 images were evaluated by our model, and simulated search times were collected. Results and Conclusion: For control target/distractor pairs, for which no asymmetry is found in humans, the model did not either exhibit spurious asymmetries. For pairs yielding asymmetry in humans, the model generally reproduced the asymmetry. With some pairs however, the model initially predicted an opposite asymmetry; careful examination of the model's internals revealed that such failure was due to luminance imbalance between target and distractor. After luminance correction, the correct asymmetries were obtained. In all asymmetry cases, the model showed significantly stronger activity in at least one feature map for the easily-found target. Our simulations hence confirm in a computational manner that asymmetries may be due to an ``added property'' in the target that is easy to detect.

Themes: Model of Bottom-Up Saliency-Based Visual Attention, Computational Modeling