Supervised Scale-Invariant Segmentation (and Detection

Abstract. The scale-invariant detection of image structure has been a topic of study within computer vision and image analysis since long. To date, Lindeberg’s scale selection method has probably been the most fruitful and successful approach to this problem. It provides a general technique to cope with the detection of structures over scale that can be successfully expressed in terms of Gaussian differential operators. Any detection or segmentation task would potentially benefit from a similar approach to deal with scale. For many of the real-world image structures of interest, however, it will often be impossible to explicitly design or handcraft an operator that is capable of detecting them in a sensitive and specific way. In this paper, we present an approach to the scale-selection problem in which the construction of the detector is driven by supervised learning techniques. The resulting classification method is designed so as to achieve scale-invariance and may be thought of as a supervised version of Lindeberg’s classical scheme