Title: Multi-scale Saliency Search in Image Analysis Location: Building 980, Room 95 (Sandia NM) Brief Abstract: Our program implements a visual saliency search mechanism, based upon the model described by Itti and Koch (*), that is purely bottom-up, that is, without any information about a specific target to find. When presented with an image, it identifies areas of interest by their conspicuity, and generates a saliency map which guides the "attention" of the model. Note that the model is purely bottom-up, it can only point out areas of interest, it cannot provide any more information about those areas. By suppressing the most salient point, attention can be shifted to the subsequent salient points, allowing the program to identify multiple areas of the image that may be of interest. This results in a kind of pre-attentive selection, that could for instance be used as a precursor to object recognition. Furthermore, it could be used to study what kinds of objects, in what kind of environment, stand out, and what kind blend into their surroundings. The algorithm relies on interesting multi-scale decomposition of data that is strongly resonant with issues inherent in multi-scale simulation. A variant of this algorithm is employed in the JPEG2000 standard for image compression as a much faster alternative to FFT. |