Object detection using a multiscale probability model

This paper presents an original scale-based image analysis method that detects objects of a given size. It relies on the comparison of two consecutive band-pass images of a multiscale analysis, so as to detect the objects that appear between the two scales. We assume that any object, present in both scales with different shapes, always contains innovative pixels, and that only non-predictable innovations are to be extracted. Thus, we compare the actual measure of innovation between scales and a model of a probability function chosen to represent the innovation between the two scales. The model describes the part of the innovation that is predictable from global statistic magnitudes. The deviations from the model are the information we wish to extract. Examples of the detection of buildings in aerial images are shown. The quality of the results is an evidence for the accuracy of the model, as well as for the efficiency of scale as a detection criterion