Self-adapting weighted operators for multiscale gradient fusion

Gradient maps are common intermediate representations in image processing, with extensive use in both classical and state-of-the-art algorithms. Most of the research on gradient map extraction has been devoted to the definition of gradient extraction operators or filters, normally by optimizing certain criteria. In this context, we find a rather limited literature in gradient map extraction using multiscale information. In this work, we develop the idea of producing a gradient map by fusing the gradient maps obtained at different scales. We first analyze the Gaussian Scale Space and the behaviour of gradients when images are projected into it; second, we propose two classes of self-adapting vector fusion operators, which are inspired by the focus-selective nature of the human visual system; third, we present a framework for multiscale boundary detection based on the use of such classes of operators for multiscale gradient fusion. We experimentally test our boundary detection framework to illustrate the validity of our vector fusion operators.