Contour Integration : Attentional Effects in a Psychophysics Task and Feature Interactions in a Computational Model

In order to achieve object recognition and image segmentation, the visual system is tasked with combining colinear and cocircular edge configurations into coherent percepts. This process is called Contour integration (CI). CI is believed to be a fundamental visual process, psychophysical experiments have shown humans to be remarkably good at integrating contours even when parts of the contours are occluded, or when a contour does not follow a smooth path. Electrophysiological studies have characterized the neural substrates of contour integration. Based on this information, modelling studies have produced algorithms to explain the functioning of putative mechanisms which give rise to CI. In this thesis, two case studies on contour integration are presented. In the first, psychophysical methods were employed to further characterize humans ability to detect contours under conditions of ambiguity. In particular, this study introduced a novel method in order to determine whether humans remarkable efficiency in detecting contours carries over to dynamic scenes. This is an important question given that scenes in nature are highly dynamic, and up to this point, most CI studies have characterized this process in static scenes. It has often been assumed that CI is a stimulus driven process which leads to pop-out percepts. Results from this study challenge these views. They indicate that humans ability to detect contours deteriorate drastically when shown extended presentations of dynamic stimuli. Furthermore, a set of sub-experiments indicates that top-down processes may play an important role in supporting contour integration under conditions of ambiguity. In the second case study, a computational model of contour integration was developed in order to account for new psychophysical findings, and further understand the mechanisms underlying these observations. Through a number of psychophysical studies, spatial frequency has been shown to be an important feature on which contours can defined and detected, and which can interact with the process of integrating oriented elements. Thus, a modulation component was added to a structurally simple model of contour integration in order to reproduce these findings. The modulation was based on the assumption that interactions of feature detectors are stronger if their preferred spatial frequencies are similar, rather than dissimilar. Extensive numerical simulations were carried out in order to understand the mechanisms leading to the mentioned psychophysical observations, and to reproduce said psychophysical results. This thesis presents contributions to the field of contour integration in two areas. In psychophysics, not only do the results from the experiments reported provide support for the emerging idea that CI may be supported by top-down process, but a significant methodological contribution was also made. A new technique to study CI was introduced. This will allow future research to characterize contour integration under new conditions. In the modeling field, a gap was bridged. To the knowledge of the author, the model presented in this thesis is the first to account for the geometrical characteristics of stimuli and the spatial frequency component of elements in the stimuli.