The contribution of global and local features to scene ‘gist’ perception

The study of scene gist perception examines how observers are able to gain an understanding of a scene within a single fixation, and the critical features of a scene that serve as cues. The general consensus within this field of research is that low-level, global scene properties related to its structure are the primary cues used in this process. Analysis of global image features is proposed to enable rapid activation of high-high level, semantic scene information, without the need to analyse finer details. However, there are some findings that suggest that more local information is both available and useful during the early stages of scene processing. The general aim of this thesis was to examine whether global scene cues are sufficient for scene gist perception, and whether scenes are represented at an abstract level early in processing. The importance of global scene cues was examined through experiments investigating how perturbing global scene structure through rotation or scrambling affects categorisation of images. The usefulness of a more specific global cue, Fourier amplitude spectrum, was also investigated by testing how equating this statistic across images affected scene categorisation. While these experiments demonstrated that global image structure generally, and Fourier amplitude in particular, are potentially useful cues, more local features still play an important role in scene categorisation. The issue of how scenes are represented during early processing was examined through the use of a Repetition Blindness (RB) paradigm. This technique was used to examine what features of scenes are important to their early representation; scenes that are similar on critical features will undergo RB. The parameters of RB for scenes were also of interest, as these had not been previously examined. Instead of RB, it was found that repetition facilitated performance, and this facilitation was based on how visually similar the two images were. There was an advantage for identical or mirror-images, but no advantage for scenes of the same basic-level category. This indicates that scenes are represented in terms of view-specific features, rather than abstract, category-general ones. Subsequent experiments revealed that changing the level at which participants were asked to categorise images influenced which features were incorporated into early representations. Category-general features facilitated superordinate-level categorisation, and scenes underwent RB when categorisation was not required. This indicates that rather than categorisation proceeding from the detection of a fixed set of features, it can instead be accomplished from a number of different cues within an image, and observers are flexible in their use of these cues. Thus how scenes are represented during early processing depends on both the properties of the stimulus, and the task an observer has been asked to perform with it