A scale-space filtering approach for visual feature extraction

Abstract

This paper presents a new integrated approach for detecting visual features which include CORNERs, ENDs, ARCs and LINEs. The effect of scale-space filtering on visual features is studied in detail as it forms the theoretical basis of our work. In this approach, the outline of the object is first extracted and it is then smoothed by scale-space filtering at different scale levels. Subsequently, the Local Extreme Curvature Points extracted from the smoothed curve and END candidates are determined to guide the termination of the filtering process. Information about the curvature of each point at the largest scale level is used to detect the different kinds of visual features. Several algorithms are proposed to determine CORNERS, ENDs, ARCs and LINEs. Experimental results show that our approach is robust to translation, rotation and scaling of the object as well as noise corruption. In addition, efficient visual features can also be successfully extracted with this approach.