Gilles Le Chenadec

Asymmetric power distribution model of wavelet subbands for texture classification

The generalized Gaussian distribution (GGD) is a well established statistical model for wavelet subband characterization used in several applications. However, it is not really suitable for eventual asymmetry of probability density functions. Therefore, in this paper we propose to exploit the asymmetric power distribution (APD) which is a more general and flexible model than the GGD. The APD parameters are estimated through the maximum-likelihood estimation. A supervised texture classification problem is proposed as an application in this work. It is based on the Bayesian framework which has led to the definition of the closed form of the corresponding Kullback-Leibler divergence considered as a similarity measure. To validate the APD model, the goodness-of-fit using the classical Kolmogorov-Smirnov test is used. Finally, classification results on four databases demonstrate the interest of the proposed approach.

Range-independent segmentation of sidescan sonar images with unsupervised SOFM algorithm (self-organizing feature maps).

The sidescan sonar records the energy of an emitted acoustical wave backscattered by the sea floor, orthogonally to the track followed. The statistical properties of the backscattered energy change with the nature of the sea floor, which allows for a segmentation of the seabed into homogeneous regions. However, the statistical description of the backscattering is not constant over the full swath of the sonar. Several parameters such as the geometry of the array or the time varying gain can be easily compensated or inverted. Making the backscattered energy independent of the grazing angle is a more difficult change, conventionally solved by considering a flat seabed and by using either Lamberts law or an empirical law estimated from the sonar data. To avoid the definition of a physical law describing the change in energy with grazing angle, the proposed algorithm divides the slant range into small stripes, where the statistics can be considered unaltered by the grazing angle variations. The starting stripe at mid sonar slant range is segmented with an unsupervised classifier based on the Kohonen algorithm SOFM (Self-Organizing Feature Maps). Then, from the knowledge acquired on the segmentation of this first stripe, the classifier adapts its segmentation to the neighboring stripes, allowing slight changes of statistics from one stripe to the other. The operation is repeated until the beginning and the end of the slant range are reached. Segmentation performances of the proposed algorithm are compared with those of conventional algorithms.

Seafloor characterization for ATR applications using the monogenic signal and the intrinsic dimensionality

In mine warfare context, environmental effects are known to degrade performances of most of automatic target recognition (ATR) processes. In this study, we consider the environment as an information that can be used to design a robust ATR process. Hence, we investigate a way to extract and exploit information about the seafloor using an isotropic analysis of sidescan sonar images based on the monogenic signal. This tool provides an orthogonal separation between energetic, geometrical and structural information of the 2D signal in a scale-space framework. It also allows to efficiently compute the continuous intrinsic dimensionality scale-space. We propose to use these last descriptors to characterize the sidescan sonar images in terms of homogeneous, anisotropic and complex areas. In each of these areas it can be expected that adapted ATR processes could be defined to outperform classical global approaches.

Seabed classification using a steerable multibeam echo sounder

The new steerable 3D multi-beam echo sounder system SEAPIX offers different operational modes to acquire supplementary information during the acquisition process. In this paper, we focused on 2 modes: the classical MBES mode and the longitudinal mode. The consistency of the data acquired from these two modes was verified through the evaluation of statistical features (mean, standard deviation, Kullback-Leibler divergence). Besides, the longitudinal mode also proposes a new data type: pixel-based data, which offers a complete backscattering profile for each pixel. A comparison of classification performance was performed with a standard Bayesian method. The pixel-based data clearly outperformed the classical modes data (93.36% vs. 50.65% using a level-based method and 81.5% using a profile-based method).

Detection of MLO in sand ripple seafloor using the monogenic signal and intrinsic dimensionality

In mine warfare context, performances of automatic target recognition (ATR) processes are known to highly depend on the underwater environment. In this study, we focus on the detection of mine-like-objects (MLO) in sand ripple seafloors. This particular seafloor type degrades detector performance by decreasing the detection probability and increasing the number of false alarms. To tackle this issue, we propose a method based on the monogenic signal and the concept of intrinsic dimensionality to characterize ripple seafloors, and then identifying geometrical anomalies caused by the presence of a MLO. This environmental information about the ripple field is taken into consideration to automatically create a map in which the potential armful object is set clearly visible to facilitate its detection. To validate and assess the performance of the proposed detection algorithm, a study of real synthetic aperture sonar (SAS) images containing various mine-like targets is undertaken.