PhD Oral Exam - John O. Brand

Abstract

It is well documented that observers are able to accurately extract the semantic information from natural scenes in 120 msec (Thorpe, Fize, & Merlot, 1996). This rapid categorization ability is often cited as evidence that the information that is required to categorize a scene originates from low-level visual information. Information related to an image’s spatial scales (Oliva & Schyns, 1997; Schyns & Oliva, 1994), phase (Joubert, Rousselet, Fabre-Thorpe, & Fize, 2009; Loschky et al., 2007, 2010; Loschky & Larson, 2008), overall summary statistics (Evans & Treisman, 2005), and colour (Castelhano & Henderson, 2008; 2005; Loschky & Simons, 2004; Oliva & Schyns, 2000) have all been shown to provide diagnostic scene information. The experiments reported in this dissertation were designed to address the overarching question of how the visual system selects diagnostic scene information? It addressed this question by examining the hypothesis that visual attention facilitates the selection of diagnostic scene information.

In order to investigate this hypothesis, the present work was divided into two main manuscripts. Manuscript 1 is presented in Chapter 2 and includes four experiments that were designed to investigate if attending to global and local levels of a scene facilitate categorization based on a scene’s coarse and fine information, respectively. This hypothesis was explored by asking observers to classify hybrid images. A hybrid image combines the coarse information (conveyed by an image’s low spatial frequencies) of one image (e.g., a city) and the fine information (conveyed by an image’s high spatial frequencies) of a second image (e.g., a highway). Experiments 1 and 2 showed that although observers could classify hybrid images based on both fine and coarse information (i.e., as either a city or a highway scene; Experiment 1), observers preferred to base categorization on coarse content (Experiment 2). Experiment 3 demonstrated that categorization based on coarse content was facilitated when observers were prompted to attend globally to scenes compared to when they were prompted to attend locally.

Experiment 4 demonstrated that this global facilitation effect was due, in part, to the facilitation of a hybrid’s low spatial frequencies.

Manuscript 2 is presented in Chapter 3 and contains four experiments that investigated the hypothesis that distributed attention facilitates the extraction of a scene’s overall summary statistics, which in turn, facilitates the ability to rapidly categorize scenes (Evans & Treisman, 2005). This hypothesis was investigated by examining whether manipulations of attention affected scene categorization in the same fashion as the extraction of overall summary statistics. Experiment 1 replicated the result that extraction of a scene’s summary statistics is more compatible with distributed attention than focused attention (Chong & Treisman, 2005). Experiments 2 and 4 extended this finding by demonstrating that superordinate level categorization of both animals (e.g., detect the presence [or absence] of an animal, Experiment 2), and natural scenes (e.g., was the scene natural? Experiment 4), were more compatible with distributed than focused attention. However, Experiment 3 showed that there was no difference between the effects of distributed and focused attention on basic level categorization (e.g., was this a beach scene?).

Together, the findings of this thesis demonstrated that visual attention is important in the rapid categorization of a natural scene, by facilitating the selection of diagnostic scene information that is necessary to classify a scene category.