Small object classification performance of high-resolution imaging sonars as a function of image resolution

Abstract

Several results from an investigation of the relationship between the classification performance of high-resolution imaging sonars and image resolution and image signal-to-noise ratio (SNR) are presented. The primary goal of this investigation has been to develop a capability to accurately estimate the classification performance of various high resolution imaging sonars used for minehunting. An additional goal has been to develop a baseline measure of classification capability that can enable more accurate evaluations of the relative performance capabilities of developmental CAD/CAC algorithms. The investigation is being conducted using synthetic sonar images, due to a severe lack of real ground-truthed image data sets with directly comparable object types, multiple resolutions, and multiple calibrated SNR values, lmage data sets containing an equal number of synthetic images with equal range and cross-range resolutions of 1, 3, 6, and 9-inches were created. The individual image data sets for these four resolutions include the same mine and minelike objects a t the same ranges and orientations to facilitate a direct performance comparison. The image backgrounds are Rayleigh distributed and the image SNR values range from 3 to 15 dB. The classification performance results were obtained using one of the advanced computer-aided detection and classification (CAC) algorithms that are currently in the process of transitioning to several US. Navy minehunting systems.Standard Receiver Operating Characteristic (ROC) curves for the joint probability of detection and classification and probability of false alarm as a function of "effective SNR" are presented. INTRODUCTlONThe Navy has a long-term requirement to be able to quickly, reliably, and safely detect and identify underwater objects in order to conserve resources, not endanger assets, and to save time when minehunting. Unfortunately, in most littoral environments, the detection process produces a large number of minelike objects that must be identified. Minelike objects have an acoustical appearance that is consistent with that of a mine. Ideally, identification would directly follow detection, but most identification sensors are relatively short ranged (i.e., < I O meters).Therefore, to reduce the number of minelike objects that must be inspected by a short-range identification sensor; a long-range classification process must follow detection. This classification process is normally performed using a high resolution imaging sonar (e.g., side-looking sonar). This has led to the ongoing cycle of development and demonstration of more effective, "more powerful" highresolution imaging sonars with the goal of achieving increased classification performance. However, the process of quantifying the actual classification performance as a function of the critical sonar parameters such as image resolution and signal-to-noise ratio (SNR) has been seriously neglected due to a lack of sonar image data and the associated ground-huth information (e.g., object type, percent burial, ra...

Combining Model-Based and in Situ Performance Prediction to Evaluate Detection and Classification Performance

“…Therefore, this research proposes a comprehensive model whereby the predictable sensorobject relationships are thoroughly characterized beforehand while the unpredictable environment or situation dependent components are estimated and incorporated in situ. This model is an extension of that proposed by Manning 6 .…”

Investigation of measureable parameters that correlate with automatic target recognition performance in synthetic aperture sonar

“…Model-based approaches [5] [6] [7] [8] use a model developed, a priori, to map parameters to ATR performance. For mission pre-planning, these parameters can either be provided based on previous knowledge of the operation area or estimated using a physics-based sonar model [9] .…”

“…For in-stride performance estimation the parameters would be measured in situ from the sonar data. In the first model [5] , Pd and Pc, are considered separately. Pc is modeled as a function of the sensor resolution, the object size, the image SNR and the background pixel statistics [6] .…”

“…The synthetic sonar images have various single-point and two-dimensional correlation parameters that were produced using the procedures outlined in the references [19] [12] [20] [21] . The sonar image pixel model used here assumes the K-distributed sonar image pixel intensity ( , ), (2), is the result of the product of an exponential distribution, representing the noise-like speckle present in sonar images, and a gamma distribution, (5), that represents the variation in the seabed texture (also called the sonar cross section) [21] . Correlation is introduced into the model by a correlated gamma random variable that is synthesized by the procedure outlined in the references [19] [20] [21] .…”

A real-time small immobile object recognition system using wavelet moment invariants

“…Therefore, it is particularly significant to detect and recognize the immobile threatening small targets. In recent years, many researchers have proposed lots of methods to solve the above problems, but most of them focused on the effectiveness and accuracy of the algorithm proposed in their papers [2][3][4]. However, to rapidly and accurately detect and identify the hostile objects is a more important factor that affects the success of modern high-tech war.…”