Benoit Zerr

Intensity-Based Block Matching Algorithm for Mosaicing Sonar Images

Registering sonar images to correctly describe seafloors and explain wide geological or biological phenomena is often achieved manually requiring significant human resources. This paper proposes an automatic intensity-based registration algorithm that relies on the optimization of a new similarity measure (SM), within a multiresolution block matching framework. Indeed, several SMs have been evaluated and ranked on real sidescan sonar data to determine the most relevant intensity dependencies between images for matching purposes. Correlation ratio (CR) and mutual information (MI) are then selected and because of their complementary behaviors, merged in a new SM (MI&CR), which performs better than CR or MI alone, to determine robust matching blocks between images. Thus, the proposed two-step registration algorithm uses MI&CR to match two sonar images: a single rigid translation globally matches the images, then a field of locally applied translations is computed for adjusting the final registration to remaining local distortions. Actual processing time can then be tuned according to the required registration accuracy. Due to a survey standard operating mode, only same-survey overlapping images are considered as candidates for matching. Moreover, building mosaics from registered images assumes a flat sea bottom as no global elevation information is provided by sidescan sonar images.

A neural network architecture for automatic extraction of oceanographic features in satellite remote sensing imagery

The authors discuss an approach for automatic feature detection and sensor fusion in remote sensing imagery using a combination of neural network architecture and Dempster-Shafer theory of evidence. Deterministic or idealized shapes are used to characterize surface signatures of oceanic and atmospherically fronts manifested in satellite remote sensing imagery. Raw satellite images are processed through a bank of radial basis function (RBF) neural networks trained on idealized shapes. The final classification results from the fusion of the outputs of the separate RBF. The fusion mechanism is based on Dempster-Shafer (DS) evidential reasoning theory. The approach is initially tested for detecting different features on a single sensor, and then is extended to classifying features observed in multiple sensors.

Acoustic data fusion devoted to underwater vegetation mapping

This paper presents research tasks conducted by SEMANTIC TS, in collaboration with GESMA, aimed to develop a mapping method for underwater vegetation lying on seabed. First stage is to develop a method for detecting and characterizing vegetation on the seabed using the acoustic response from a conventional single beam echo sounder. This new method is then operated simultaneously with multibeam sonar producing micro‐relief information and side scan sonar providing gray scale levels associated with bottom reflectivity. Then fusion of these three data is processed. We show efficiency of these multisensor data fusion concept to get very precise seabed vegetation mapping in a way reducing truth control (video and diving investigations). Sensors and method accuracy allow obtaining, like in biomedical field, real 3D scan pictures of seabed vegetation. This study is first applied to posidonia and cymodocea, which play a key role in Mediterraneans echosystem. Then, extension of the method is investigated to addre...

Sonar image enhancements for improved detection of sea mines

In this paper, five methods for enhancing sonar images prior to automatic detection of sea mines are investigated. Two of the methods have previously been published in connection with detection system and serve as reference. The three new enhancement approaches are a variance stabilizing log transform, nonlinear filtering, and pixel averaging for speckle reduction. The effect of the enhancement step is tested by using the full processing chain i.e. enhancement, detection and thresholding to determine the number of detections and false alarms. Substituting different enhancement algorithms in the processing chain gives a precise measure of the performance of the enhancement stage. The test is performed using a sonar image database with images ranging from very simple to very complex. The result of the comparison indicates that the new enhancement approaches improve the detection performance.

Bracketing the solutions of an ordinary differential equation with uncertain initial conditions

In this paper, we present a new method for bracketing (i.e., characterizing from inside and from outside) all solutions of an ordinary differential equation in the case where the initial time is inside an interval and the initial state is inside a box. The principle of the approach is to cast the problem into bracketing the largest positive invariant set which is included inside a given set X. Although there exists an efficient algorithm to solve this problem when X is bounded, we need to adapt it to deal with cases where X is unbounded.

Model based classification of mine-like objects in sidescan sonar using the highlight information

This paper presents a model-based approach to perform underwater target classification. Very high resolution imaging sonar has increased the opportunities to use highlight information contained in the target acoustic signature whereas underwater target classification is still mainly based on the analysis of geometrical properties of the acoustic shadows. Supervised classifiers generally use experimental or simulated samples of target acoustic signature in the training stage but when the testing set is different from the training set the performance can be altered. Here the classification method consists in comparing the A-scan of the detected target with a set of simulated A-scans generated by our Sonar Image Simulator (SIS) in the same operational conditions. The used simulator relies on acoustical ray tracing techniques and takes into account complicated underwater physical process to simulate an accurate time response of underwater targets (A-scan). Practically the classifier is made of a cascade of matched filters. Each is built by simulating the A-scan for a given object in a given orientation (and /or for a given size). The resulting scores can be used to rank likelihood of belonging to object classes. The result is flexible and gives a percentage match for each class. With this approach the training set can be extended increasingly to improve classification when classes are strongly correlated. This classification process is assessed on a few real sidescan sonar data. These first results are finally discussed and further work is deduced to improve the general classification task.

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.

Detection of objects on the sea bottom using backscattering characteristics dependent on the observation point

Sector-scanning sonar systems image the sea bottom to detect objects that can be distinguished from the background structure of the sea bottom. In current systems, images are displayed and discarded as new image data become available, In this paper, a method for improving sonar detection by utilizing all images in a sequence is investigated. The proposed method requires that sonar data are acquired with a sector-scanning sonar in a side-looking configuration. It is demonstrated that these data can be used to detect observation-point-dependent changes in sea-bottom backscattering characteristics. These changes provide additional cues for discrimination that can improve the detection of objects on the sea bottom. Results of applying the method to experimental data are presented.

An Interval Approach to Compute Invariant Sets

This paper proposes an original interval-based method to compute an outer approximation of all invariant sets (such as limit cycles) of a continuous-time nonlinear dynamic system, which are included inside a prior set of the state space. Contrary to all other existing approaches, our method has the following properties: first, it is guaranteed (a solution cannot be lost); second, it is applicable to a large class of systems without any specific assumption such as the knowledge of a Lyapunov function or any partial linearity; and third, there is no need to integrate the system.

An interval approach to solve an initial value problem

This paper proposes an original guaranteed interval-based method to solve an Initial Value Problem (IVP) for ordinary differential equations (ODE). Our method uses the geometrical properties of the vector field given by the ODE and a state space discretization to compute an enclosure of the trajectories set that verifies the IVP problem.