Target Recognition in Synthetic Aperture and High Resolution Side-scan Sonar

The accurate detection and identification of underwater targets continues as a major issue, despite, or perhaps as a result of, the promise of higher resolution underwater imaging systems, including synthetic aperture sonar, high frequency sidescan and high resolution cameras. Numerous techniques have been proposed for Computer Aided Detection to detect all possible mine-like objects, and Computer Aided Classification to classify whether the detected object is a target or not. The majority of existing techniques have concentrated on the analysis of the acoustic shadow casted by the objects. With the development of ultra high-resolution systems, new techniques inherited from the computer vision community can now be applied, taking into account the acoustic echo and shadow. This paper will briefly review traditional techniques and present new possible solutions to the problem of detection and classification in synthetic aperture sonar. 1 Target Detection and Classification With the recent advances in autonomous underwater vehicle (AUV) technology for mine-countermeasures (MCM) the need has arisen for on-board automated techniques for object identification from sonar data. Research carried out in developing MCM tools is generally split into Computer Aided Detection (CAD)[1-4] to detect all possible mine-like objects, and Computer Aided Classification (CAC) models [6-11] to classify whether the detected object is a mine or not. A common approach is to compare a set of extracted features from the mine-like object (MLO) [6,3] to a set of predetermined training data. The system is trained using a set of ground-truthed data before being run on the unknown “test” data. These approaches require a large amount of training data often unavailable as real experiments are very expensive and are often limited in scale. This can lead to over-fitting of the data when small datasets are used and limits the generalisation capabilities of such approaches [3]. Fusing the results from multiple classifiers using different feature sets and classification techniques can improve performance [11]. In general, these systems offer a black-box solution to the problem, and it can be difficult to analyse why a particular result is obtained. Obtaining information other than the basic mine or not-mine label is usually referred to as object identification. Information such as the shape and dimension of the mine can allow the mine type to be determined and can help detail how best to neutralise the threat. Man-made objects such as mines generally have regular shapes and so leave regular shaped shadows in sidescan images. The shape of these shadows can be used to identify the objects by extracting relevant features from the shadows and comparing these to known training data [8]. The non-linear nature of the shadow-formation process ensures a shadow normalisation step is required for these approaches to be widely applicable. This paper will briefly discuss and compare two possible solutions to this problem of trained systems. The first is a model based technique which uses a three stage process to detect and classify the targets, but which requires no training. The second solution is to employ simulation and augmented reality simulators to permit the generation of training data. This in turn enables the simulation of targets within real sonar data, allowing the training data to be tailored to the specific sonar system or the operating environment. With large amount of training data available and the improvement in resolution gained from synthetic aperture or very high frequency imaging (acoutics cameras), modern computer vision techniques can be used for detection and classification. This will be briefly demonstrated for each scenario. 2. High Resolution Imaging Typical sidescan systems used for mine countermeasures applications can provide a resolution of between 3cm and 10cm depending on operating range and frequency. New very high frequency systems, such as the Marine-Sonics 2.4MHz system offers resolution of 1cm but is limited to operating ranges of less than 6m. Although such systems offer increased potential for improved classification, their use for detection of targets is limited as a result of their limited range. The first commercially produced Synthetic Aperture Sonar (SAS) systems are now available, bringing the promise of higher resolution surveys at long ranges. In the area of Mine Countermeasures this offers the capability to detect mines at longer ranges and provide higher resolution images of targets for subsequent classification. However, little work has been done to investigate the automatic detection and classification of mines from SAS imagery, since previous research in this area concentrated on developing the sensor rather than the automated processing of the images, since it was always assumed that their higher resolution would simplify this task. However although providing higher resolution, the images contain a significant level of speckle due to the construction of the image. Filtering methods have been employed, but these can degrade either the shadow or the highlight. The shadow can also be altered by the SAS processing limiting the applicability of well established detection techniques in sidescan imagery. 3 Overview of Model Based Systems Model based systems use strong prior information of the image formation process to guide the detection or identification of a target. We present here a model based system tailored for side-scan sonar imagery with potential applications to SAS data. The model based technique [4,10] for the detection and classification of mine-like objects (MLOs) presented here has been developed to overcome the problems of trained systems with a three-stage process which is summarised in figure 1. The first stage is the detection of mine like objects (MLOs) within the sidescan sonar image. This stage uses a Markov Random Field (MRF) model to directly segment the image into regions of object highlight, shadow and background [4]. Unlike many previous models for object detection this requires no training. The structure of the MRF model also integrates specific priors that take into account the characteristics of sidescan data. Simple information such as the fact that a target is generally displayed as an area of object highlight followed by shadow, and that the highlight tends to form small clusters surrounded by background pixels can readily be integrated. Figure 1: Overview of Detection and Classification System The detected targets from this first stage are then passed to stage two, where the highlight and shadow regions of the detected objects are extracted using a Cooperating Statistical Snake technique (CSS)[4]. The CSS model approximates the background as three homogenous regions – object highlight, background and shadow and so uses two statistical snakes, one to segment the highlight and one the shadow. The a priori information between the highlight and shadow is used to constrain the movement of the snakes so as to achieve accurate segmentation results regardless of the seabed type involved. The CSS model can also be used to eliminate false alarms from the detection stage. Areas can be indicated as object highlight and false alarms produced, especially from more complex regions such as sand ripples. If the CSS model is applied to these regions, since they do not have the expected characteristics of MLOs, the shadow snake will expand in an uncontrolled manner. If the snakes expand beyond mine-like dimensions the detection can be identified as a false alarm and removed. After a detected MLO has passed through both the Detection and CSS modules, stage 3 of the system will classify the object. To do this, the system has the extracted shadow and object-highlight region of the MLO, as well as some simple information extracted from the navigation data, such as height of the sonar above the seabed and the range to the object at the time of ensonification. Although the highlight is generally not considered for classification purposes since it is generally unpredictable and difficult to model, basic information can be extracted from it and used together with the information from the shadow region. The model represents possible mine-like shapes (cylinder, sphere, truncated-cone) using parametric models which allow a sonar simulator to generate the resulting shadow region from such an object. Each shadow region is specific to the particular parameters of each shape and is generated under the same sonar conditions as the MLO was detected. As the model searches through the different parameter options, the resulting synthetic shadows are compared to the real MLO shadow to find the best match for each considered class. This section of the model is the most computationally intensive and a detailed overview of this section is shown in Figure 2. Figure 2: Determination of best synthetic shadow match to real shadow Once this has been completed, the degree of match and the shape parameters used to obtain it are output to the classifier to define a class membership function. These membership functions are entered into a classifier which identifies the belief that the object belongs to a particular class. As indicated in figure 1, the classification can also be extended to include multiple views [10] if the target has been seen from multiple angles. 3.1 Results with Sidescan Table 1 illustrates the output of the classification system for 4 example sidescan images extracted from Edgetech DF1000 and Klein 5500 sonar data. This illustrates the successful classification of these targets. Belief Sidescan Images Bel(cyl) Bel(sph) Bel(con) Bel(clut) Cylinder 0.71 0 0 0.2 Sphere 0 0.833 0 0.083 Cone 0 0.303 0.45 0.045 Clutter 0.42 0 0 0.46