Research on the automatic analysis of sonar images has focused on classical, i.e. non deep learning based, approaches for a long time. Over the past 15 years, however, the application of deep learning in this research field has constantly grown. This paper gives a broad overview of past and current research involving deep learning for feature extraction, classification, detection and segmentation of sidescan and synthetic aperture sonar imagery. Most research in this field has been directed towards the investigation of convolutional neural networks (CNN) for feature extraction and classification tasks, with the result that even small CNNs with up to four layers outperform conventional methods. The purpose of this work is twofold. On one hand, due to the quick development of deep learning it serves as an introduction for researchers, either just starting their work in this specific field or working on classical methods for the past years, and helps them to learn about the recent achievements. On the other hand, our main goal is to guide further research in this field by identifying main research gaps to bridge. We propose to leverage the research in this field by combining available data into an open source dataset as well as carrying out comparative studies on developed deep learning methods.

This paper proposes a new method for segmenting and classifying seamines on Synthetic Aperture Sonar (SAS) side scan images. The method uses an active contours approach and superellipse a-priori knowledge to segment the image in object, object-shadow and background areas. In contrast to other methods using superellipse constraints, the shape prior is incorporated directly into the segmentation process. This kind of segmentation has the advantage that afterwards the extracted superellipse parameters that describe the object and the object-shadow can directly be used as feature for a classification - this work is hence also of potential interest for general object recognition tasks in other application domains. Several different perspectives of implementing this idea into a suitable algorithm are introduced and compared with each other. Thus, for the evaluation of each method the extracted superellipse features are used for a support vector machine classification. An one against all confusion matrix is generated on a test data set. This result is compared to a related state of the art algorithm. It is shown that our new method is able to correctly classify 170 of 210 objects in a very challenging real world data set and that it yields significant better results than the state of the art comparison.

This paper deals with several original contributions to an automatic target recognition (ATR) system, which is applied to underwater mine classification. The contributions concentrate on feature selection and object classification. First, a sophisticated filter method is designed for the feature selection. This filter method utilizes a novel feature relevance measure, the composite relevance measure (CRM). Feature relevance measures in the literature (e.g., mutual information and relief weight) evaluate the features only with respect to certain aspects. The CRM is a combination of several measures so that it is able to provide a more comprehensive assessment of the features. Both linear and nonlinear combinations of these measures are taken into account. A wide range of classifiers is able to provide satisfactory classification results by using the features selected according to the CRM. Second, in the step of object classification, an ensemble learning scheme in the framework of the Dempster–Shafer theory is introduced to fuse the results obtained by different classifiers. This fusion can improve the classification performance. We propose a reasonable construction of the basic belief assignment (BBA). The BBA considers both the reliability of the classifiers and the support of individual classifiers provided to the hypotheses about the types of test objects. Finally, this ATR system is applied to real synthetic aperture sonar imagery to evaluate its performance.

In this paper we deal with an unsupervised segmentation approach for images given by a synthetic aperture sonar (SAS). The images with objects are segmented into highlight, background and shadow. Since the shape features are extracted from these segmented images, correctness and precision of the segmentation are highly required. We improve the expectation-maximization (EM) methods of Sanjay-Gopal et al. by using the gamma mixture model. Moreover an intermediate step (I-step) based on Dempster-Shafer theory (DST) is introduced between the E- and M-steps of the EM to consider the pixel spatial dependency. Finally, numerical tests are carried out on both synthetic images and SAS images. The results are compared to iterative conditional mode (ICM) and diffused EM (DEM). Our approach provides segmentations with less false alarms and better shape preservation.

This paper presents a new filter method for feature selection, which maximizes a hybrid relevance measure using a sequential forward searching scheme (mHRM-SFS). An individual relevance measure provides the classification information only in a certain aspect. Thus, we choose a modified Relief weight, mutual information, and the information entropy to formalize a hybrid relevance measure (HRM) to identify the most characterizing features. Because of its efficiency, a sequential forward searching scheme is used for maximizing the HRM. The resulting feature selection obtained by mHRM-SFS is appropriate to serve as an input for various classifiers. The mHRM-SFS can also determine the cardinality of the feature selection automatically while choosing the optimal features. Finally, the mHRM-SFS is applied to select features of underwater objects for the classification purpose. The selected features are tested by different classifiers. The classification results are compared to those of 4 existing feature selection methods.

This paper proposes a new level set segmentation method which is guided by a parametric prior force based on Lamé curves. The use of prior knowledge is advantageous in order to improve the segmentation results in terms of matching expected object types because one can in general state that for different applications some shapes are more likely than others. By avoiding complex shape training processes the level set idea is extended with a parametric prior shape which forces the level set evolution to propagate towards the desired objects by not overpowering the image properties. Also a cross talk evolution is discussed for ternary images to handle correlations between adjacent or correlated objects.

This paper addresses the calculation of moments of complex Wishart and complex inverse Wishart distributed random matrices. Complex Wishart and complex inverse Wishart distributed random matrices are used in applications like radar, sonar, or seismology in order to model the statistical properties of complex sample covariance matrices and complex inverse sample covariance matrices, respectively. The moments of these random matrices are often needed e.g. in studies of asymptotic properties of parameter estimates. This paper gives a derivation of the probability density function of complex inverse Wishart distributed random matrices. Furthermore, strategies are outlined for the calculation of the moments of complex Wishart and complex inverse Wishart distributed matrices.

We address the problem of estimating the components of superimposed exponentially decaying signals. Usual estimation techniques, e.g. least squares or eigenvalue and eigenvector based methods, are not adequate for exponentionally decaying fluorescence processes due to their rather simple signal modelling. Therefore, we introduce a better suited parametric model by exploiting the statistical properties of the exponentially decaying emission of fluorescence photons (time dependent Poisson statistics). Using this model maximum likelihood estimates for the fluorescence intensity spectrum and the decay parameters are derived. The performance of the maximum likelihood estimates is compared with the least squares estimates by means of simulations and real data experiments. The results indicate the superiority of the maximum likelihood estimates.

The problem of source location estimation in the presence of partly unknown noise fields is addressed. A novel two-stage procedure is developed which combines the conditional maximum-likelihood estimate and the conditional marginal maximum-likelihood estimate for the signal parameters and the noise parameters, respectively. The strong consistency and asymptotic normality of the conditional maximum-likelihood estimates for location parameters and noise parameters in the case of not necessarily normal and independent distributed observations are proved. Alternatively, different least squares criteria fitting a parametric model of the spectral density matrix to a nonparametric consistent estimate of the spectral density matrix are investigated and their asymptotic behaviors are mentioned briefly. Results of numerical experiments are presented to show the performance of the different estimates.>

The direction-of-arrival estimation of signal wavefronts in the presence of unknown noise fields is investigated. Generalizations of known criteria for both conditional and nonconditional maximum-likelihood estimates are developed. Numerical calculations show that the usual Gauss-Newton iteration for conditional maximum-likelihood estimates cannot give good results. Therefore, a related, relatively simple two-step least-squares estimate is constructed. Results of numerical experiments are presented and indicate that the two-step estimate has approximately the same power as the least-squares estimate using the exact noise correlation structure.>

The authors present an adaptive method to detect transient signals of unknown waveform in colored Gaussian noise. The signal is modeled as the impulse response of a rational transfer function. After a prewhitening of the observed signal, the order of the rational transfer function is estimated by a method based on the covariance matrix perturbation and the coefficients by a maximum-likelihood estimator with constraint. The detection of the signal is performed by means of the generalized likelihood ratio test. The method is compared to an adaptive detection scheme using an autoregressive (AR) model of the signal and to the conventional energy detector. Theoretical detection probabilities are compared to the experimental ones obtained by Monte Carlo simulations.>