A SALIENCY MAP MODEL EXPLAINS THE EFFECTS OF RANDOM VARIATIONS ALONG IRRELEVANT DIMENSIONS IN TEXTURE SEGREGATION AND VISUAL SEARCH

Li Zhaoping, Presented at Annual meeting of Society of Neuroscience, Orlando Florida USA, Nov. 2-7, 2002.

Abstract: Consider visual search or texture segmentation tasks when image items are oriented lines, and that target/distractors and texture regions are defined by the line orientations. It has been shown (Snowden, J. Exp. Psychol. 1998) that random feature variations along irrelavant dimensions such as color significantly degrade performances of texture segregation but has no or only small effects on visual search. To explain the observations, I apply a model of a saliency map in V1 (Li, TICS 2002), and relate the ease of a task to the salience of the target or the texture border. The salience of an item increases with its evoked V1 response relative to the background responses to the whole image. The response of a V1 cell is determined not only by the input contrast within its classical receptive field, but also by the contextual stimuli via intra-cortical interactions. The model shows that random variations of the stimuli along irrelevant dimensions such as color increase the responses of the color tuned cells, as well as the response inhomogeneity, to the background stimuli. Such an increase can reduce the salience or pop out strengths of the target and the texture border whose salience values are largely determined by cells tuned to orientation --- the relevant feature dimension. The model shows that the salience value of the target is affected much less than that of the texture border. This is because the target does not experience iso-orientation suppression which affects the texture border elements. Consequently the response to the target, unlike the response to the texture border, is too high to be (partly) submerged by the background responses. Supported by: The Gatsby Foundation