Learning to Detect Contours with Dynamic Programming Snakes

Contour detection is an important and challenging task in computer vision, with many applications in the analysis of natural scenes and biomedical images. Although there are many general approaches to contour detection, achieving good performance in any given application often requires considerable hand-tuning of algorithm parameters, optimization criteria, or pre-processing of the images themselves. We propose a novel framework for contour detection that combines learning of a probabilistic classifier with dynamic programming-based contour optimization. On test images, we find that our system is able to learn to detect specific types of contours in images, often from just a single example contour. After learning, the system can be used to speed up interactive contour detection, requiring the user only to click once on each target object, or it can be used to automatically detect all contours of the same type in an image or set of images.