Rank-Order Filters for FOPEN Target Detection - IEEE Xplore

0 downloads 0 Views 139KB Size Report
Rank-Order Filters for FOPEN Target Detection. Atindra K. Mitra, Senior Member, IEEE, Thomas L. Lewis, Member, IEEE, and Arnab K. Shaw, Member, IEEE.
IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 2, FEBRUARY 2004

93

Rank-Order Filters for FOPEN Target Detection Atindra K. Mitra, Senior Member, IEEE, Thomas L. Lewis, Member, IEEE, and Arnab K. Shaw, Member, IEEE

Abstract—Issues associated with the radar detection of military targets that are concealed or camouflaged by forest clutter are described. The specific sensor platform can be categorized as an ultrawideband (UWB) foliage penetration (FOPEN) synthetic aperture radar (SAR). The discussion illustrates the fact that many contemporary approaches to FOPEN target detection are computationally intensive and/or require the implementation of elaborate training procedures. Alternative approaches, based the application of a set of simple rank-order filters (alternately known as order statistical or L filters), are presented. Initial results indicate impressive performance levels (in terms of probability of detection as a function of false-alarm rate) with respect to baseline constant false-alarm rate computations. A number of avenues for future investigations are cited. Index Terms—Automatic target detection, foliage penetration, image processing algorithms, synthetic aperture radar (SAR), ultrawideband (UWB) radar.

I. INTRODUCTION

I

N RECENT YEARS, military experts within the Department of Defense (DoD) have expressed an interest in developing airborne capability of detecting targets that are concealed by foliage. In terms of specific DoD focus areas, this problem falls under the general category of counter camouflage, concealment, and deception (CC&D). This particular CC&D problem has been termed generically as “FOPEN” (foliage penetration). FOPEN addresses the issue of detecting military vehicles, when these vehicles are relocated from open terrain into forested regions to avoid detection. (Note that the term FOPEN is commonly used both as a generic term that describes the foliage penetration problem, and it is also the name of an existing DOD Advanced Technology Demonstration program called FOPEN SAR.) Possible additional nondefense applications for FOPEN technology might include drug trade intervention, emergency search and rescue, as well as some environmental applications. In terms of FOPEN physical phenomenology, early research efforts focused on basic studies such as characterizing foliage attenuation versus frequency. Properties of backscattered radar waveforms, or radar cross sections, in the X-band region have been studied for many years, and many existing airborne military radars make use of this region of the electromagnetic spectrum. For example, flying a typical X-band synthetic aperture

Manuscript received July 11, 2001; revised July 24, 2002. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Peyman Milanfar. A. K. Mitra is with the Air Force Research Laboratory, Wright-Patterson AFB, OH 45433-7333 USA (e-mail: [email protected]). T. L. Lewis was with Sverdrup Technology, Dayton, OH. He is now with the Air Force Research Laboratory, Wright-Patterson AFB, OH 45433-7333 USA (e-mail: [email protected]). A. K. Shaw is with the Electrical Engineering Department, Wright State University, Dayton, OH 45435 USA (e-mail: [email protected]). Digital Object Identifier 10.1109/LSP.2003.819856

radar (SAR) (i.e., high-resolution imaging radar) over a forest generates a map of the upper canopy layer. In order to achieve sufficient level of foliage penetration and to generate radar scattering from targets that are concealed by forest clutter, studies have shown that it is necessary to lower the frequency of the radar waveform to at least the UHF region of the electromagnetic spectrum. Operating the radar in the UHF and VHF regions poses a number of unique challenges with regard to designing radar hardware [2], developing high-quality SAR image formation algorithms [3], and developing high-performance FOPEN target detection algorithms. This letter addresses the issue of high-performance FOPEN target detection. Fig. 1 shows a conceptual block diagram of a typical FOPEN target detector as proposed in a recent paper [4], where further details on the individual blocks in the diagram can be found in [5]. This letter deals with the preliminary detector stage, which is physically located in the airborne portion of the overall detector, and it consists of a screener that feeds into a Bayesian neural network (BNN) [4], followed by a target/clutter segmenter. As outlined in [4] and [5], the BNN functions as a discriminator, and it requires an elaborate training process. Extensive training requirements, in turn, bring issues to the forefront with regard to the robustness of the BNN under a variety of testing and training operating scenarios that may require training over one (conveniently accessible) region and testing over another region [1]. In the next section, a number of simple rank-order filters are proposed as possible alternatives to the BNN-based preliminary airborne detector. These filters are investigated with the goal of providing computationally efficient albeit high-performance alternatives to the current FOPEN target detection algorithms. By replacing the preliminary detection stage with a rank-order filter, the training requirements and computational complexity of the airborne preliminary stage can be relaxed considerably. II. RANK-ORDER FILTERS FOPEN target detection usually occurs in “impulsive clutter” environment where target detection algorithms based solely on constant false-alarm rate (CFAR)-type computations typically do not achieve a sufficiently low false-alarm rate. In terms of phenomenology, the “impulse clutter” might be attributed to the FOPEN SAR image formation process where long coherent integration intervals [3] are required in order to achieve good image resolution at long UHF and VHF wavelengths. These long integration intervals tend to generate many instances of bright clutter points due to integration over relatively aspect-independent ground-to-tree-trunk bounce response. It is well known that classical linear filters are not appropriate for removing “impulsive” or “spiky” noise. Linear filters also tend to blur edges that can destroy target-size infor-

1070-9908/04$20.00 © 2004 IEEE

94

IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 2, FEBRUARY 2004

Fig. 1.

Block diagram of a contemporary approach to UHF FOPEN target detection at lockheed martin (a commercial developer) [4].

mation in the UWB image. To address these limitations, a variety of signal-dependent rank-order-type filters have emerged in the image processing community in the past two decades [6], [7]. These nonlinear filters have demonstrated good noise suppression characteristics in many environments with “spiky” noise characteristics. A number of filters that fall within this class of rank-order filters that suppress “impulse noise” and preserve important details within an image are summarized in [6] and [7]. For the purposes of this preliminary investigation, several rank-order filters were adopted for FOPEN target detection application. The order statistical filtering can be described by considering the following general filter function:

Fig. 2.

C. Modified Nearest Neighbor (MNN) for center otherwise

(1) where, is the result of sorting in ascending image window. Choice order the data from an of and depends on the size of a target to be detected. is the number of pixels used for averaging, and denotes the filter coefficients. and depend on the choice of particular rank order filter, as defined next. A. Alpha-Trimmed

In this case,

(2)

.

for median otherwise.

center

(4) is the pixel value at the center of the image in this case is the number of pixels that satisfies center .

D. Median

for otherwise.

(5)

.

E. CFAR Detector [8]

int

B. Modified Alpha-Trimmed (MTM)

In this case, median

where center window, and center

In this case, for int otherwise.

Outer and inner box configuration used in CFAR.

median

(3) is the number of pixels that satisfies, median .

This is the most commonly used approach, and it will be used for baseline comparisons with the rank-order filters. Further details on CFAR as applied to SAR target detection may be found in [8]. Briefly, as illustrated in Fig. 2 and (6), the CFAR score is the ratio of the average pixel value in the inner box divided by the average pixel value in the outer box. The size of the inner , in general) is selected to conform to the approxbox ( imate size of the object of interest. The outer box is intended to provide an estimate for the average value for localized clutter, and the guard band is meant for separation of target and clutter. In order to obtain an accurate estimate for the clutter average, the lower bound on the size of the guard ring is selected in order

MITRA et al.: RANK-ORDER FILTERS FOR FOPEN TARGET DETECTION

95

Fig. 3. ROC curves using rank-order filter for FOPEN sample image. CFAR result is plotted for baseline comparison.

to avoid contamination from possible target samples form the inner band. The upper bound on the guard ring is selected in order to constrain the clutter samples to remain within the localized area of interest and to avoid contamination from neighboring target samples [8]. In general, the guard band and outer . For box widths would usually be a fraction of example, based on the known target geometry and pixel resolution, if a particular target size is expected to be 12 10, then an inner box size of 12 10 is chosen, and the widths of the guard band and outer box can be 3 and 5, respectively CFAR

average of pixels in inner box average of pixels in outer box

(6)

The following two rank-order filters were formulated at Air Force Research Laboratory (AFRL) particularly to address the FOPEN target detection problem. The primary motivation is to reduce the effect of heavy tailed impulsive clutter on the performance of CFAR. The inner/outer box configuration of CFAR depicted by Fig. 2 is used in these two cases also, although averaging is done over either the inner or outer box. F. Inner-Sigma Filter

Fig. 4. Section of original image in (a) versus inner-sigma filtered image in (b).

Fig. 3 demonstrates the detection results based on receiver operating characteristics (ROC) curves for all of the above filters for one sample FOPEN image. Fig. 4 is a display of the center section of the original sample image alongside the inner-sigma filtered image. Observation of the inner-sigma filtered image shows more contrast between the “blob-like” FOPEN targets and the clutter background in comparison to the original image with “spiky” impulse clutter. III. INTERPRETATION OF RESULTS AND FUTURE DIRECTIONS

for mean (inner box) otherwise.

std inner box

(7)

In this case, is the number of pixels that satisfy mean(outer box) std(outer box) . G. Outer-Sigma Filter for mean (outer box) otherwise.

std outer box

In this case, is the number of pixels that satisfy mean(outer box) std(outer box)

(8)

A set of rank-order filters were formulated and developed for FOPEN target detection applications. As indicated by the results in Fig. 3, a number of these filters are capable of superior target detection performance with respect to a CFAR baseline. The inner-sigma filter, formulated at AFRL, is generating good target detection performance with respect to many of the other filters. This filter was the result of an initial attempt to develop a simple filter that “pseudoadaptively” estimates and removes the effect of the “clutter tail” that is used as a rank-order filter parameter. On the other hand, the outer-sigma filter, also formulated at the AFRL, is generating lower levels of performance with respect to the other filters. This may indicate that it might

96

be difficult to estimate the parameters of “impulse-type” clutter from a narrow outer (CFAR-like) box. Increasing the size of this outer box may have adverse effects such as generating samples from neighboring targets within the outer “clutter sample” box. A number of directions can be cited for purposes of future investigations as follows. • As more FOPEN data is collected in the future, compare the performance of simple rank-order filters with the performance of sophisticated FOPEN algorithms from commercial developers. Possibly consider integrating rank-order filters into commercial developer’s algorithm suite as risk mitigation measure. • Investigate “clutter-adaptive rank-order filtering techniques.” This might involve extending the inner-sigma filter technique to more sophisticated approaches of estimating the “clutter tail.” For example, objective functions that measure the deviation of the extreme regions of the clutter histograms from an ideal Gaussian distribution can be formulated and used within a clutter-adaptive rank-order filtering scheme. • Consider developing and applying optimal rank-order filter algorithms that estimate more appropriate values for the filter coefficients. • Consider extending UHF FOPEN target detection results to rank-order filters for VHF FOPEN target detection. • Pursue basic and advanced clutter modeling investigations in order to gain better understanding for FOPEN clutter. For example, one approach might be to investigate the am-

IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 2, FEBRUARY 2004

plitude levels of the impulse clutter versus the length of the SAR coherent integration interval/angle. A number of other advanced approaches can be applied toward a basic study to model and understand FOPEN clutter at a deeper level. The resulting information might be applied toward developing more sophisticated rank-order filters. • As more FOPEN data are collected in the future, measure the robustness of the performance of rank-order filters as a function of different FOPEN operating environments. REFERENCES [1] J. G. Fleischman, S. Ayasli, E. M. Adams, and D. R. Gosselin, “Foliage attenuation and backscatter analysis of SAR imagery,” IEEE Trans. Aerosp. Electron. Syst., vol. 32, pp. 135–144, Jan. 1996. [2] M. Davis, P. Tomlinson, and P. Maloney, “Technical challenges in ultrawideband radar development for target detection and terrain mapping,” Proc. IEEE 1999 National Radar Conf., pp. 1–6, Apr. 1999. [3] A. Mitra, J. Suriano, and A. Kerrick, “Image formation computations for a low frequency foliage penetrating radar,” in Proc. IMAGE 2000 Conf., July 10–14, 2000, pp. 234–41. [4] D. MacDonald, J. Isenman, and J. Roman, “Radar detection of hidden targets,” Proc. IEEE 1997 National Aerospace and Electronics Conf., vol. 2, pp. 846–855, 1997. [5] E. E. Hilbert and C. F. Chang, “Bayesian neural network ATR for multifeature SAR,” in Proc. SPIE Int. Symp. Optical Engineering in Aerospace Sensing, Orlando, FL, Apr. 4–9, 1994, pp. 344–355. [6] I. Pitas and A. N. Venetsanopoulos, “Order statistics in digital image processing,” Proc. IEEE, vol. 80, pp. 1893–1921, Dec. 1992. [7] , “Nonlinear digital signal/image processing: Overview and perspectives,” in IEEE Winter Workshop on Nonlinear Digital Signal Processing, 1993, pp. T_1.1–T_1.8. [8] C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Images. Norwell, MA: Artech House, 1998, pp. 279–282.