A Computational-Geometry Approach to Digital Image Contour Extraction

Abstract. We present a simple method based on computational-geometry for ex-tracting contours from digital images. Unlike traditional image processing meth-ods, our proposed method first extracts a set of oriented feature points from the input images, then applies a sequence of geometric techniques, including cluster-ing, linking, and simplification, to find contours among these points. Extensive experimental results on synthetic and natural images show that our method can effectively extract contours from both clean and noisy images. Experiments on the Berkeley Segmentation Dataset also show that our proposed computational-geometry method can be linked with any state-of-the-art pixel-based contour ex-traction algorithm to remove noise and close gaps without severely dropping the contour accuracy. Moreover, contours extracted by our method have a much more compact representation than contours obtained by traditional pixel-based meth-ods. Such a compact representation allows more efficient extraction of shape fea-tures in subsequent computer vision and pattern recognition tasks