Susanne Wenzel

Detection of repeated structures in facade images

We present a method for detecting repeated structures, which is applied on facade images for describing the regularity of their windows. Our approach finds and explicitly represents repetitive structures and thus gives initial representation of facades. No explicit notion of a window is used; thus, the method also appears to be able to identify other manmade structures, e.g., paths with regular tiles. A method for detection of dominant symmetries is adapted for detection of multiply repeated structures. A compact description of the repetitions is derived from the detected translations in the image by a heuristic search method and the criterion of the minimum description length.

Finding Poly-Curves of Straight Line and Ellipse Segments in Images Segmentierung von Pixelketten in Geraden- und Ellipsenelemente

Simplification of given polygons has attracted many researchers. Especially, finding circular and elliptical structures in images is relevant in many applications. Given pixel chains from edge detection, this paper proposes a method to segment them into straight line and ellipse segments. We propose an adaption of Douglas-Peucker’s polygon simplification algorithm using circle segments instead of straight line segments and partition the sequence of points instead the sequence of edges. It is robust and decreases the complexity of given polygons better than the original algorithm. In a second step, we further simplify the poly-curve by merging neighbouring segments to straight line and ellipse segments. Merging is based on the evaluation of variation of entropy for proposed geometric models, which turns out as a combination of hypothesis testing and model selection. We demonstrate the results of circlePeucker as well as merging on several images of scenes with significant circular structures and compare them with the method of PATRAUCEAN et al. (2012). Zusammenfassung: . ie tion runder und elliptischer Strukturen ist relevant für viele Anwendungen. Die Reduktion der Komplexität gegebener Polygone ist für sich ein interessantes Forschungsthema. Diese Arbeit stellt ein Verfahren zur Segmentierung von Pixelketten einer Kantendetektion in Geradenund Ellipsensegmente vor. Der erste Schritt besteht in einer Adaption des Douglas-Peucker Algorithmus, in der Kreise anstelle von Geraden zur Partitionierung verwendet werden und die Punktstatt der Kantensequenz partitioniert wird. Das Verfahren ist robust und reduziert die Komplexität der gegebenen Polygone stärker als der originale Algorithmus. In einem zweiten Schritt vereinfachen wir diese Vorsegmentierung durch das Verschmelzen benachbarter Segmente zu Geradenund Ellipsensegmenten und stützen uns dabei auf die Entropieänderung. Wir zeigen die Ergebnisse der Vorsegmentierung als auch der folgenden Vereinfachung anhand verschiedener Bilder von Szenen, die signifikante kreisförmige Strukturen aufweisen und vergleichen sie mit dem Algorithmus von PATRAUCEAN et al. (2012).

Discriminative archetypal self-taught learning for multispectral landcover classification

Self-taught learning (STL) has become a promising paradigm to exploit unlabeled data for classification. The most commonly used approach to self-taught learning is sparse representation, in which it is assumed that each sample can be represented by a weighted linear combination of elements of a unlabeled dictionary. This paper proposes discriminative archetypal self-taught learning for the application of landcover classification, in which unlabeled discriminative archetypal samples are selected to build a powerful dictionary. Our main contribution is to present an approach which utilizes reversible jump Markov chain Monte Carlo method to jointly determine the best set of archetypes and the number of elements to build the dictionary. Experiments are conducted using synthetic data, a multi-spectral Landsat 7 image of a study area in the Ukraine and the Zurich benchmark data set comprising 20 multispectral Quickbird images. Our results confirm that the proposed approach can learn discriminative features for classification and show better classification results compared to self-taught learning with the original feature representation and compared to randomly initialized archetypal dictionaries.