Purchase this article with an account.
Takeshi Suzuki, Yannik T. H. Schelske, Tandra Ghose; Dissociating Semantic and Pragmatic Information in Eye Movement Data for Image Processing Tasks. Journal of Vision 2014;14(10):90. doi: https://doi.org/10.1167/14.10.90.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Eye-movements recorded during image processing tasks are influenced by the image content (semantic-information) and the image processing task (pragmatic-information). To analyze the image processing task it is necessary to separate these two. We empirically test a method to transform an image into variants that contain less semantic-information while preserving task relevant features, such as color impression, spatial color correlation and luminance. Analysis of eye-movements for a global contrast adjustment task demonstrates the applicability of this method. Images were taken from two semantic categories, namely landscape and macro images. These were further divided into two subcategories each, namely landscape images with or without water and macro images containing human or nonhuman focal objects. This allowed us to assess at which semantic level (main categories represent a coarse semantic level, subcategories a fine semantic level) this semantic-information reduction method can still be applied. The images from these categories were transformed, using this method, into three variants containing different amounts of semantic-information. Subjects performed global contrast adjustment task on these images, blocked such that each subject saw all images but only in one of the three semantic-information variants. Subjects chose a similar global contrast for images independent of the variant they saw, but fixation distribution and frequency were significantly different between the variants. Eye-movement patterns show that despite high similarity between subcategory images, our semantic-information reduction technique preserved the features that elicit subcategory specific and task relevant eye-movements. Our interpretation is that this method does not remove task relevant pragmatic-information even at fine semantic level, shown by similar performances in the global contrast adjustment task for all image variants. The benefit of our method is that eye-movements are not confounded by influences of semantic-information present in the original images.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only