Presented at the International Society for Optical Engineering (SPIE 2000), San Jose, CA, 2000. Using a Model of the Human Visual System to Identify and Enhance Object Contours in Natural Images John A. Black, Jr. * , Sethuraman Panchanathan Visual Computing and Communication Lab Department of Computer Science and Engineering Arizona State University, Tempe, AZ 85287-5406 ABSTRACT Segmentation of natural images depends on the ability to identify continuous contours that define the boundaries between objects. However, in many natural images (especially those captured in environments where the illumination is largely ambient) continuous contours can be difficult to identify. In spite of this, the human visual system efficiently perceives the contours along the boundaries of occluding objects. In fact, optical illusions, such as the Kanizsa triangle, demonstrate that the human visual system can “see” object boundaries even when spatial intensity contrasts are totally absent from an image. In searching for the mechanism that generates these “subjective contours” neurological researchers have found that the 2D image on the retina is mapped onto Layer 4 of the primary visual cortex (V1) and that there are lateral connections within the 6 layers of V1 that might subserve contour completion. This paper builds on a previous model of the early visual system (including the retina, the LGN and the simple cells of V1) by adding lateral interconnections to demonstrate how these interconnections might provide contour completion. Images are presented to show how this model enhances the detection of continuous spatial contours, thus contributing to the segmentation of natural images. Keywords: contour extraction, contour completion, human visual system, contour interpolation, subjective contours, primary visual cortex 1. INTRODUCTION 1.0 Background One simple method for extracting contours from 2D images is to draw boundaries between pixels with differing intensity levels. This can be done most conveniently by thresholding the image. A single threshold value can be used to produce a binary image with white and black areas, corresponding to light and dark areas in the original image. With carefully configured, evenly-distributed lighting, and with adequate contrast between the foreground objects and the background, such a thresholding method can be useful for simple machine vision tasks, such as verifying the high-contrast silhouette of flat mechanical components as they pass under a video camera on a white conveyer belt. However, natural images are rarely captured in a tightly-controlled lighting environment, and the luminance contrast between foreground objects and the background is often very poor. Problems as simple as uneven lighting due to reflections from adjacent objects make it very difficult to choose the proper threshold value. A threshold value that is best suited for extracting useful contours from one region of an image might be very ill-suited for extracting useful contours from another region. In general, the problem is not a lack of contours, but the fact that the contours in such a thresholded image might have no particular significance. Meanwhile, other luminance contours that are significant, but don’t happen to straddle the arbitrary threshold value, are eliminated from the thresholded image. Surface textures can also be a significant problem. With uneven lighting from one portion of the image to another, it might be impossible to avoid high levels of visual noise in some regions of the image where a textured surface’s average brightness passes through the chosen threshold value. While modified thresholding techniques (such as localized threshold values) can be used, it is difficult to design a pixel-level contour extraction method that avoids the extraction of either too many or too few contours. ––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– * Correspondence Email: panch@asu.edu; Telephone: 480-966-5936; Fax: 480-968-3446