Contour detection using a multi-scale approach and surround inhibition

In this thesis an algorithm for contour detection in image processing is described. The goal is to get to a contour detector that mimics the human perception of contour edges. For this purpose a performance measure is introduced, which compares the result of a contour detector with a hand drawn extended groundtruth image. In the thesis a contour detector is gradually developed. Topics dealt with are convolution, noise reduction, gradient computation, edge thinning using non-maximal suppression and binarization. These techniques are used in a well-known edge detector, the Canny edge detector. These topics are followed by a chapter about multiscaling and a method to suppress the strength of edges originating from texture known as surround inhibition. The performance using a single and multi-scale contour detection approach, both with and without surround inhibition, is evaluated using the performance measure. Multi-scaling proves to give no absolute performance gain, but decreases the performance spread and often the median increases. The perfonnance gain of surround inhibition depends much on the signal-to-noise ratio. If texture and object contours are present at the same scale and the amount of texture contours is high, there can be a considerable increase.If one looks at the output images with highest performance created by the contour detection methods described here, one notices that the object(s) present are recognizable as the object(s) shown in the input images. Unfortunately, if the input image is for example a picture of an animal or human, parts of the outline and important facial characteristics (like eyes and ears) are often missing in these output images. A human would complete the contour information using his or her experience. The contour detection algorithms described here lack this ability.