Data sparsity is a well-known limitation in the sonar domain. This limitation is a problem when applying data-intensive techniques from the computer vision community, such as deep learning models for detection and classification. One way of extending a sonar dataset is to use simulated sonar images however, these often have the drawback of looking non-realistic when compared to domain data. To overcome the data-sparsity problem as well as for generating realistic-looking sonar images, we introduce a pipeline where the possibilities and limitations of applying cycleGAN to enhance simulated forward-looking sonar images are explored. The results show improved classification performance when training a classifier on enhanced-simulated images compared to training on solely simulated images.

Imaging and mapping sonars such as forward-looking sonars (FLS) and side-scan sonars (SSS) are sensors frequently used onboard autonomous underwater vehicles. To acquire information from around the vehicle, it is desirable for these sonar systems to insonify a large area; thus, the sonar transmit beampattern should have a wide field of view. In this work, we study the problem of the optimization of wide transmission beampatterns. We consider the conventional phased-array beampattern design problem where all array elements transmit an identical waveform. The complex weight vector is adjusted to create the desired beampattern shape. In our experiments, we consider wide transmission beampatterns (>= 20 degrees) with uniform output power. In this paper, we introduce a new iterative-convex optimization method for narrowband linear phased arrays and compare it to existing approaches for convex and concave-convex optimization. In the iterative-convex method, the phase of the weight parameters is allowed to be complex as in disciplined convex-concave programming (DCCP). Comparing the iterative-convex optimization method and DCCP to the standard convex optimization, we see that the former methods archive optimized beampatterns closer to the desired beampatterns. Furthermore, for the same number of iterations, the proposed iterative-convex method achieves optimized beampatterns, which are closer to the desired beampattern than the beampatterns achieved by optimization with DCCP.

Robust feature correspondences between 2D sonar imagery are important for perception tasks in the underwater domain such as 3D reconstruction but involve open challenges, in particular, low-resolution as well as the fact that object appearance is view-dependent. Although sonars in the MHz range would allow for higher resolution imagery, in this paper we focus on scenarios with a lower frequency kHz sensor, in which the longer visual range is gained at the sacrifice of image resolution. To this end, we first propose to solve the correspondence task using the PatchMatch algorithm for the first time in sonar imagery, and then propose a method for feature extraction based on IC. We then compare the proposed methods against conventional methods from computer vision. We evaluate our method on data from a lake experiment with objects captured with an FLS sensor. Our results show that the proposed combination of IC together with PatchMatch is well-suited for point feature extraction and correspondence in sonar imagery. Further, we also evaluate the different methods for point correspondence with a 3D object reconstruction task.

Several beamforming techniques can be used to enhance the resolution of sonar images. Beamforming techniques can be divided into two types: data independent beamforming such as the delay-sum-beamformer, and data-dependent methods known as adaptive beamformers. Adaptive beamformers can often achieve higher resolution, but are more sensitive to errors. Several signals are processed from several consecutive pings. The signals are added coherently to achieve the same effect as having a longer array in synthetic aperture sonar (SAS). In general it can be said that a longer array gives a higher image resolution. SAS processing typically requires high navigation accuracy, and physical array-overlap between pings. This restriction on displacement between pings limits the area coverage rate for the vehicle carrying the SAS. We investigate the possibility to enhance sonar images from one ping measurements in this paper. This is done by using state-of-the art techniques from Image-to-Image translation, namely the conditional generative adversarial network (cGAN) Pix2Pix. The cGAN learns a mapping from an input to output image as well as a loss function to train the mapping. We test our concept by training a cGAN on simulated data, going from a short array (low resolution) to a longer array (high resolution). The method is evaluated using measured SAS-data collected by Saab with the experimental platform Sapphires in freshwater Lake Vättern.

The sonar image resolution is classically limited by the sonar array dimensions. There are several techniques to enhance the resolution; most common is the synthetic aperture sonar (SAS) technique where several pings are added coherently to achieve a longer array and thereby higher cross range resolution. This leads to high requirements on navigation accuracy, but the different autofocus techniques in general also require collecting overlapping data. This limits the acquisition speed whencovering a specific area. We investigate the possibility to enhance the resolution in images processed from one ping measurementin this paperusing compressive sensing methods. A model consisting of isotropic point scatterers is used for the imaged target. The point scatterer amplitudes are frequency and angle independent. We assume only direct paths between the scatterers and the transmitter/receiver in theinverse problemformulation. The solution to this system of equations turns out to be naturally sparse, i.e., relatively few scatterers are required to describe the measured signal.The sparsity means that L1 optimization and methods from compressive sensing (CS) can be used to solve the inverse problem efficiently. We use the basis pursuit denoise algorithm (BPDN) as implemented in the SPGL1 package to solve the optimization problem.We present results based on CS on measurements collected at Saab. The measurements are collected using the experimental platform Sapphires in freshwater Lake Vättern. Images processed using classical back projection algorithms are compared tosonar images with enhanced resolution using CS, with a 10 times improvement in cross range resolution.

Forward Looking Sonars (FLS) are a typical choiceof sonar for autonomous underwater vehicles. They are mostoften the main sensor for obstacle avoidance and can be usedfor monitoring, homing, following and docking as well. Thosetasks require discrimination between noise and various classes ofobjects in the sonar images. Robust recognition of sonar data stillremains a problem, but if solved it would enable more autonomyfor underwater vehicles providing more reliable informationabout the surroundings to aid decision making. Recent advancesin image recognition using Deep Learning methods have beenrapid. While image recognition with Deep Learning is known torequire large amounts of labeled data, there are data-efficientlearning methods using generic features learned by a networkpre-trained on data from a different domain. This enables usto work with much smaller domain-specific datasets, makingthe method interesting to explore for sonar object recognitionwith limited amounts of training data. We have developed aConvolutional Neural Network (CNN) based classifier for FLS-images and compared its performance to classification usingclassical methods and hand-crafted features.