Abstract
This project presents Log-Polar Boundary Extraction (L-PoBE) as a model of human visual system in a boundary extraction task. Given an input image, the model selects and connects the set of pixels belonging to an object boundary to form a closed shape. In human visual processing, boundary extraction consists of two parts: the edge-based boundary contour system that forms boundaries of shapes, and the feature contour system that fills in the color information to the surface of objects. Our model is an extension of Kwon et al.’s (2016) model, which was designed for noisy monochromatic line drawing images. Our new model can be applied to noisy colored images, such as that of a parrot in its natural habitat. We show that consistent color at either region alone, at the foreground or background, facilitates the boundary extraction process. Our synthetic stimuli consisted of geometric shapes with colored noise added to only one region, either inside or outside of the shape. Noise patterns were inserted randomly with a particular density, thickness, and color variation. L-PoBE extracts color information from a small region near the detected edges, includes the differences between color as an additional term in the cost function, and solves a global optimization problem using Dijkstra shortest path algorithm in the log-polar space. To encode directionality, each edge in the image is represented by two nodes in the graph, allowing for all 4 permutations of pairwise interpolations. In the current version of the model, the start edge is selected as the longest detected edge belonging to the boundary of the target shape. We expect, based on Kwon et al.’s results, that this initialization will not be needed. The model-extracted boundaries are shown to be robust across a wide variety of noise parameters and fixation positions.