Target detection in clutter for sonar imagery
Abstract
This thesis is concerned with the analysis of side-looking sonar images, and specif-
ically with the identification of the types of seabed that are present in such images,
and with the detection of man-made objects in such images. Side-looking sonar
images are, broadly speaking, the result of the physical interaction between acous-
tic waves and the bottom of the sea. Because of this interaction, the types of
seabed appear as textured areas in side-looking sonar images. The texture descrip-
tors commonly used in the field of sonar imagery fail at accurately identifying the
types of seabed because the types of seabed, hence the textures, are extremely
variable. In this thesis, we did not use the traditional texture descriptors to identify
the types of seabed. We rather used scattering operators which recently appeared
in the field of signal and image processing. We assessed how well the types of
seabed are identified through two inference algorithms, one based on affine spaces,
and the other based on the concept of similarity by composition. This thesis is
also concerned with the detection of man-made objects in side-looking sonar im-
ages. An object detector may be described as a method which, when applied to
a certain number of sonar images, produces a set of detections. Some of these
are true positives, and correspond to real objects. Others are false positives, and
do not correspond to real objects. The present object detectors suffer from a
high false positive rate in complex environments, that is to say, complex types
of seabed. The hypothesis we will follow is that it is possible to reduce the
number of false positives through a characterisation of the similarity between the
detections and the seabed, the false positives being by nature part of the seabed.
We will use scattering operators to represent the detections and the same two
inference algorithms to quantify how similar the detections are to the seabed.