Guiding early visual processing with qualitative scale and region information

Abstract

Many methods in computer vision and image analysis implicitly assume that the problems of scale detection and initial segmentation have already been solved. One example is in edge detection, where the selection of step size for the gradient computations leads to a trade-off problem. A small step size gives a small truncation error, but the noise sensitivity might be severe. Conversely, a large step size will in general reduce the noise sensitivity, but at the cost of an increased truncation error. In the worst case a slope of interest can be missed and meaningless results obtained, if the difference quotient approximating the gradient is formed over a wider distance than the size of the object in the image. The problem originates from a basic scale problem, namely that the issue of scale must be considered when selecting a mask size for computing spatial derivatives.