Man-made objects vs natural objects in SAS data

During the last decade a lot of effort has been made on the ATR (automatic target recognition) field. The main goal is the automatic detection / classification / identification of objects of interest. The traditional approach comes from the analysis of acoustic images given by sidescan or forward looking sonar systems. The main image processing methods used so far are either model based, learning based or saliency based approaches. The introduction of SAS (synthetic aperture sonar) increases drastically the resolution of the image, and by doing so the probability of detection and classification. Going back to the main goal of ATR, one can argue that the definition of an object of interest is ill-stated. Looking at the MCM (mine countermeasure) context, the models of mines are not necessarily well known and new models will come up with new shapes and new structures. The introduction of NEDs (non-ordinary explosive device) moves the problem of ATR from looking for certain types of mines to looking for man-made objects. In this paper we will discuss a generic method to describe and model natural objects and man-made objects. In our approach natural objects are represented as 3D volumes with rough fractal surfaces while man-made objects can be modeled by a 3D volume with smooth geometrical shapes. As a consequence the backscattering echo from a man-made object can be approximated by a finite number of scatterers. The SAS image formation results from the integration of the echo over a continuous multi-view aspect. In this paper we study the phase stability over the multi-views provided by a SAS system in order to differentiate between a stable scatterer and a rough surface.