Knowledge-based methods for small object detection in SAR images

Abstract

Various techniques for detecting anomalous reflective objects that extend a few tens of pixels in a SAR image are discussed. These techniques assume some prior knowledge of the object, to make the detection process robust and minimize false alarms. The basis of the method is a CFAR technique that assumes Weibull clutter statistics. The method is then augmented using knowledge of the shadow cast by the object, change detection, and information from terrain databases. Examples using a simulated SAR image are given.

An Improved Scheme for Target Discrimination in High-Resolution SAR Images

“…1) We can combine the algorithms of feature-and knowledge-based discrimination. In practice, using the method based on knowledge independently yields poor results [27]. Yet, it is still an effective supplement to the algorithm based on feature extraction.…”

“…The second method is a context-based method [25]- [28], in which an estimation is made based on both the environment and the shadow of the target. However, execution of a complicated terrain classification algorithm to obtain information on the environment is difficult [27]. In addition, the shadow of the target is not robust enough (e.g., the shadow does not appear when the target is placed in the background with weak scattering, such as in a wide concrete road) [25], [31].…”

“…1) Among the knowledge-based methods, both changedetection [27] and context-based [25]- [28] methods for removing CFAs are far from practical application. Only the methods based on target-group knowledge are relatively easy to implement, and thus can be employed in real applications of automatic target discrimination [10], [29], [30], [32].…”

“…Such a method has been termed as change detection, which relies on image-registration techniques and sequential observations of the same area [27]. The second method is a context-based method [25]- [28], in which an estimation is made based on both the environment and the shadow of the target.…”

“…Knowledge-based target discrimination can be achieved using three methods. The first method [27] is to observe the same area at different times and to compare new images with the known images of nontargets.…”