Real-Time Detection of Object Contours using Artificial Visual Attention

S. Reinke


The main objective of this work is to identify, implement and analyze methods for generating binary edge images using artificial visual attention. The edge images should only contain high quality and informative binary edges of objects so that they can be used for contour based object recognition systems. Generally, edge detection can be divided into gradient- and learning-based approaches. Gradient-based algorithms are quite fast, but also detect uninformative edges, e.g. background edges, which are not helpful to describe an object. In contrast, learning-based approaches have a high run time but can generate high-quality edge images. Additionally, artificial visual attention (AVA) models can be used for edge detection. Most of the AVA models are inspired by human visual attention and use different concepts to generate edge images. One possibility here is to analyze the mechanism behind AVA models and the way they highlight "interesting" areas in images. In this work it will be analyzed how such mechanisms can be used for edge detection. Other AVA methods highlight regions of interest (ROI) in an image, where objects are likely to appear. It will be investigated if a reduction of run time can be achieved if the ROIs are analyzed instead of the whole image while still generating appropriate edge images.