IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
2099
Segmentation of ISAR Images of Targets Moving in Formation Sang-Hong Park, Hyo-Tae Kim, and Kyung-Tae Kim
Abstract—This paper proposes a new method to separate inverse synthetic aperture radar (ISAR) images using the rangeDoppler algorithm when multiple targets fly closely spaced in a formation. This method is composed of three steps. The first step is an initial range-Doppler imaging of the whole set of targets to obtain a “bulk” image. In this step, range profiles are aligned using a new cost function, which is the sum of the amplitudes of the pixels lying along a polynomial that models the trajectory of the bulk image. To reduce the probability of incorrectly aligning high-amplitude pixels from different targets, pixel amplitudes are converted to binary form: A number η of pixels with the greatest amplitudes are identified in the bulk range profile, and their amplitudes are converted to one; the amplitudes of the remaining pixels are converted to zero. This process gives each of the η pixels the same weight on the cost function. Then, phase adjustment is used to coarsely separate targets in a 2-D image. The second step is the separation of the bulk image into component images using a window of the target size. The third step is a second range-Doppler imaging, in which each ISAR image is enhanced using range alignment and phase adjustment. Simulations using three targets composed of point scattering centers prove that the proposed method can effectively segment three targets flying in a formation. Index Terms—Formation flight, image segmentation, inverse synthetic aperture radar (ISAR), particle swarm optimization (PSO), polynomial.
I. I NTRODUCTION
I
NVERSE synthetic aperture radar (ISAR) imaging is a technique to generate a 2-D high-resolution image of a target [1]. An ISAR image of the target can be generated by synthesizing radar signals obtained from various observation angles. Due to its efficient 2-D features, ISAR has many military applications [2]. Currently, the main limitation to ISAR imaging is motion compensation; this is because targets may have complicated motion components, which can degrade the resolution of the image. Algorithms such as range-Doppler [3] and time– frequency analysis [4] can compensate for the motion component of a single target, but, when multiple targets occur in a single radar beam, it is difficult to image each target sepa-
Manuscript received May 7, 2009; revised July 12, 2009. First published November 3, 2009; current version published March 24, 2010. This work was supported by the Brain Korea 21 Project. S.-H. Park and H.-T. Kim are with the Department of Electronic and Electrical Engineering, Pohang University of Science and Technology (POSTECH), Pohang 790-784, Korea (e-mail:
[email protected];
[email protected]). K.-T. Kim is with the Department of Electronic Engineering, Yeungnam University, Gyeongsan 712-749, Korea (e-mail:
[email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2009.2033266
rately. Algorithms have been proposed to solve this problem, but none is fully satisfactory. One proposed method assumes that the number of targets is known [5], and another requires that the range migrations of all targets be equal [4]. Both of these methods are impractical in real cases. Other conventional methods also have limitations [6], [7]. This paper proposes a method to separate ISAR images when aircraft targets fly in a formation. This method is composed of three steps: 1) an initial range-Doppler imaging of the whole set of targets to obtain a “bulk” image; 2) separation of the bulk image into component images using an appropriate segmentation algorithm; and 3) a second range-Doppler imaging, in which each ISAR image is enhanced using range alignment and phase adjustment. The first step is the main subject of this paper. In this step, a general range-Doppler algorithm is applied to coarsely separate targets in the 2-D range-Doppler domain. In carrying out the translational motion compensation that is composed of range alignment and phase adjustment, we show that currently used cost functions in range alignment (e.g., correlation and entropy) can yield poor results because high scattering centers may be aligned even if they do not belong to the same range bin. To solve this problem, this paper proposes a new range alignment method which aligns range profiles using a polynomial that best represents the trajectory. We use particle swarm optimization (PSO) to estimate the parameters of the polynomial. In calculating the cost function, we construct a new image by selecting the highest η amplitudes in each range profile, then convert their amplitudes to one and the others to zero. After this conversion, each of the η pixels has the same weight on the cost function. We also introduce a segmented alignment algorithm to align the range profiles if the polynomial does not fit the flight trajectory. Processes used in this step position targets separately in the 2-D range-Doppler profile. The second step is the segmentation of the ISAR image of each target separately positioned in the 2-D range-Doppler domain. In this step, we utilized a simple method which groups pixels that can be accommodated in a window of the target size, belonging to a target. In this process, we calculate the average pixel power, then convert all pixels to one if their values are greater than this average and to zero otherwise. This conversion of pixels to one or zero is distinct from the process described in the previous paragraph. We search the image systematically from the start of the image to find the first element having a value of one, then extract the pixels that occur within the window centered on that element. Then, we calculate the center of mass (COM) of the extracted data and draw a new window centered on the calculated COM. COM is recalculated for this
0196-2892/$26.00 © 2009 IEEE Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
2100
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
window; then, a new window is drawn centered on this value of COM. This process is repeated until the number of pixels within the window stops increasing. At this point, the window is considered to contain the target completely. The third step is the enhancement of each image. In this step, range alignment and phase adjustment are repeated again after converting each segmented range-Doppler image into a rangetime domain using inverse fast Fourier transforming (FFT) in each range bin. Then, FFT is performed in each range bin to derive the ISAR image of each target. In simulations using three targets composed of point scattering centers, the proposed method successfully separates the ISAR images of the targets. II. S IGNAL M ODEL AND P ROPOSED M ETHOD A. Signal Model and Range-Doppler Algorithm For the radar signal, we assume a monostatic chirp waveform (1) because it is widely used for high range resolution 2 t j2π f0 t+ Bt 2τ × rect r(t) = A0 e (1) τ where r(t) is a transmitted signal at time t, A0 is the amplitude of the signal, f0 is the start frequency, B is the bandwidth, τ is the pulse duration, and rect is a function whose value is one for t − (τ /2) ≤ t ≤ t + (τ /2) and zero otherwise. Then, the received signal reflected from a scattering center becomes B(t−d0 )2 j2π f0 (t−d0 )+ t − d0 2τ × rect r(t) = Ae (2) τ where A is the amplitude of the reflected signal and d0 is the time delay between the radar and the scatterer. The two-way attenuation of the emitted signal is not considered in (2) for simplicity. In the case of multiple targets, the received signal g(t) is B(t−dk,l )2 K L j2π f0 (t−dk,l )+ t−dk,l 2τ Ak,l e × rect g(t) = τ k=1 l=1 (3) where K is the number of targets, L is the number of scattering centers, Ak,l is the amplitude of scattering center l of target k, and dk,l is the time delay between the radar and the scattering center. The value of dk,l is calculated using plane wave approximation, in which the distance to a scattering center is that projected onto the radar line-of-sight vector. The reflected signal is matched filtered to obtain range profiles at a certain aspect angle; then, after translational motion compensation by range alignment and phase adjustment, FFT is applied to each range bin to resolve scattering centers in the cross-range direction (azimuth FFT). Translational motion compensation is composed of two steps: range alignment which aligns range profiles and phase adjustment which removes phase errors caused by the direct shift of each range profile. In range alignment, methods that exploit the similarity of the envelopes of the range profiles by using cost functions such as correlation [3] and entropy [8] are
widely used. The 1-D entropy cost function H is efficient and relatively unaffected by noise. It is defined as follows: Np −1
HGm ,Gm+1 = −
G(k, n) ln G(σ, n)
0
where G(τ, n) =
|Gm (n)| + |Gm+1 (n − σ)| N −1
.
(4)
(|Gm (n)| + |Gm+1 (n − σ)|)
0
Gm (n) and Gm+1 (n) are the range profiles m and (m + 1) of range bin n, respectively, Np is the total number of range bins, and σ is an integer shift that is used to align the profiles. According to this criterion, σ that minimizes the 1-D entropy is the shift that best aligns range profile (m + 1) with range profile m. Two improved alignment methods have been proposed to better align the range profiles. One [9] utilizes the entropy of the average range profile as a global quality measure to be immune to the accumulation of misalignment in [8], but the method in [9] requires an exhaustive search operation. The other proposed method [10] avoids this exhaustive search by utilizing the Shannon entropy of the average range profile. However, both [9] and [10] are based on [8] because they both use the entropy as a cost function. In phase adjustment, methods such as the maximum-contrast method [11] and the minimum-entropy method [12] are used because they can be applied even if no information is available about the motion of the target. B. Limitations of the Existing Alignment Methods When Multiple Targets Occur Existing range alignment results have limitations because they were derived for a single target. Alignment using entropy (4) is effective in imaging a single target because neighboring range profiles are highly correlated. However, this method cannot always be applied when several closely spaced aircraft targets fly in a formation with almost identical velocity and acceleration. Generally, these aircraft are located in a single radar beam, and, as a result, scattering centers from different targets can occur in the same range bin, causing constructive and destructive interference. This interference can cause range profiles derived for several aspect angles to be highly uncorrelated. In this case, the range alignment performance of the entropy or other correlation-based cost functions becomes questionable. In this paper, we only consider the entropy cost function (4) because this method is more robust than others [8] and because correlation yields similar results to the entropy method because the correlation method uses sum after multiplication, which is similar to (4), which uses multiplication after sum. The entropy cost function (4) is inadequate as an indicator of the alignment when multiple targets exist (Fig. 1). In the case of a single target in which two range profiles are highly correlated and are separated by one range bin [Fig. 1(a)], the two range profiles are highly correlated and, therefore, are correctly aligned using the shift derived by minimizing the entropy. However, in a multiple-target case [Fig. 1(b)], where
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
PARK et al.: SEGMENTATION OF ISAR IMAGES OF TARGETS MOVING IN FORMATION
Fig. 1.
2101
Alignment results using entropy. (a) Single target. (b) Two targets.
the values in range bins 3 and 5 suddenly increased to six and seven due to the influence from other targets, the alignment is not correct: Range bins 3 and 5 in range profile 2 should be aligned with bins 4 and 6 in profile 1. This misalignment occurs because G(σ, n) in (4), which determines the entropy, is the sum of the two shifted range profiles, so the entropy decreases as this sum increases. Therefore, range bins with high amplitudes contribute more to the entropy than do bins with low amplitudes. As a result, range alignment causes high-valued range bins to be aligned. In the single-target case, this method
is acceptable because neighboring range profiles are similar, but, in the multiple-target case, this method is not appropriate because neighboring bins are dissimilar. Fig. 1 shows only a simple case; this problem will be more serious in the case of real targets flying in a formation. C. Proposed Range Alignment Method Because bins with large amplitudes have disproportional influence on the entropy calculation, range alignment using
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
2102
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 2. Proposed cost function. The sum of the pixels in the target region was used as the cost function.
entropy is not a good method in the multiple-target case. To solve the problem of misalignment due to high-valued range bins, we propose a new method that satisfies the following two requirements. 1) All pixels, not just those having high values, must have the same effect on the range alignment. 2) The cost function must be appropriate for the multipletarget case. To achieve 1), we construct a new image instead of using the range profile history itself. For this purpose, we select η range bins having the highest amplitudes in each range profile and change their values to one. This process makes all η pixels identical, so they have the same weight in the range profile calculation. The amplitudes in the other range bins are all changed to zero. The value of η should be chosen properly to select range bins that correspond to scattering centers. If η is too large, range bins that correspond to sidelobes can be selected, and, if it is too small, only some major range bins are selected. In general, the number of range bins in each range profile having amplitudes that are higher than 10% of the maximum value in the range profile is calculated, and η is selected between the maximum and minimum numbers in the set of range profiles. In this paper, the range of η was 40 ≤ η ≤ 60. To achieve 2), we model the shifts of the range profiles (R(t)) as a polynomial R(t) = 1+a1 t+a2 t2 +a3 t3 . . . ,
t = 0, 1, . . . , M −1
(5)
where M is the number of range profiles and ai are the fitted parameters. We round R(t) to the nearest integer. We then define the cost function as the sum of the values of the pixels that occur at the polynomial location on the new image. If the sum is large, the polynomial represents the trajectory well, and the
alignment can be accomplished easily using the fitted parameters. However, because range bins belonging to the targets in a range profile history comprise a region, not a single line, serious errors can occur if one maximum number of the integrated value is used for the alignment. Instead, shifting R(t) sequentially from the first to the last range bins and summing the pixel values on the polynomial for each shift (Fig. 2), we utilize as the cost function the sum of the η highest sums that best represents the range profile history. Therefore, the cost function yields the maximum value when the polynomial fits the trajectory well. To reduce the computation time, we can reduce the search space by using only the shifts that yield a cost function > 0 by positioning the trajectory in the target region (Fig. 2). We use PSO to find the coefficients in (5) because it is easy to implement and has proven to be efficient for several engineering problems [13]. The system is initialized with a population of random solutions (particles) that minimizes a cost function; then, the algorithm searches for local and global particle optima by changing the velocity vector of each particle. The particle dynamics which updates each particle is as follows: vi (t) = φvi (t − 1) + ρ1 (xpbest − xi (t)) + ρ2 (xgbest − xi (t)) (6) where t is the generation number, φ is the inertial weight, ρ1 = r1 c1 , ρ2 = r2 c2 , c1 , c2 > 1 and c1 + c2 < 4, and ri is drawn from a random uniform distribution from zero to one. Then, the velocity vector vi (t) in generation t is added to the position vector xi (t) to move this particle. If the polynomial defined in (5) does not represent the trajectory accurately, serious alignment errors can result. This mismatch between the trajectory and the polynomial is possible because the flight trajectory can vary significantly depending on the initial position, the direction vector, and the motion
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
PARK et al.: SEGMENTATION OF ISAR IMAGES OF TARGETS MOVING IN FORMATION
Fig. 3.
Segmented alignment.
Fig. 4.
Proposed method to segment ISAR images.
parameter. When the polynomial and the observed trajectory differ, we divide the range profile history into several segments with each pair of neighboring segments sharing one range profile (Fig. 3). Then, the proposed alignment method is applied to each segment, and neighboring range profiles are aligned using the overlapping range profile as a guide. In this procedure, more time is needed than when using one polynomial; in theory, parallel processing can be used to eliminate this increased time requirement. We do not examine this possibility in this paper. Comparing the computational requirements of the proposed methodology with the existing alignment methods, the proposed method is computationally more efficient than the methods in [8] and [9] because the cost is a simple summation of M η pixels, where M is the number of pulses. However, it is slower
2103
than the method in [10] because it utilizes a sequential search based on an iterative approach, whereas the proposed method utilizes a population-based approach. D. Proposed Segmentation Procedure The proposed method to segment ISAR images is composed of three steps: the first range-Doppler imaging, segmentation, and the second range-Doppler imaging (Fig. 4). The main subject of this paper is the first range-Doppler alignment, which utilizes the method proposed in Section II-C. This range alignment method positions targets in the 2-D rangeDoppler domain at separate locations that correspond to their range and Doppler values after the range-Doppler algorithm is
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
2104
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 5. Image segmentation using a window of the target size; Cn = COM of window n. Pixel values > average are converted to one and others to zero. Sums of the converted pixel values within windows are compared sequentially until the sum stops increasing.
executed. The existing alignment method does not position the targets separately because high-valued range bins are aligned regardless of the target to which they belong; this makes segmenting targets very difficult. After the first step, ISAR images containing several targets are thresholded using the mean pixel value I of the image, which is given by 1 I(i, j) P Q i=1 j=1 P
I=
Q
(7)
where I(i, j) is the (i, j)th pixel value of an ISAR image, P is the number of down-range bins, and Q is the number of crossrange bins. The purpose of this step is to remove clutter effects caused by mutual interference among different targets. The second step is the segmentation of each target. In this paper, we segmented the targets by specifying a window with dimensions corresponding to the size of the target, then moving this window from the start to the end of the image and grouping image pixels that can be accommodated within the window. A new binary image was constructed for this purpose by converting pixels that have power that is higher than the average (7) to one and the others to zero. The image was scanned, starting from the first element in the first row, until the first pixel with a value of one was located. Using this pixel as the center of the window, we extracted the pixels within the window. Then, COM of the extracted data was calculated as COM = (Cx , Cy ) =
Q P
Ib (i, j) × (i, j)
i=1 j=1
where Ib =
Ib Q P Ib (i, j) i=1 j=1
(8)
and Ib is the binary image windowed; P and Q are the same as those in (7). A new window was created centered on the calculated COM (8), and COM was recalculated within the new window. This process was repeated until the number of converted pixels within the window stopped increasing (Fig. 5). More sophisticated segmentation algorithms are available [14]–[16]; improvement of this step is left for future work. The final step is the enhancement of each segmented ISAR image. Because interference can blur the segmented image, we enhance the image by converting the range-Doppler image into a range-time image. Then, range alignment and phase adjustment are performed again. After this alignment, only one target exists in each segmented image, so (4) is used to align the range profiles. After phase adjustment and FFT in each range bin, enhanced ISAR images can be derived. III. S IMULATION R ESULTS In this simulation, we assume three targets separated along the x-axis on the same xy plane (i.e., altitude). Two types of targets are used (Fig. 6); these were modeled using the 3-D CAD data of Russian Su-35 and American F-14 fighters (www.3dcadbrowser.com). The Su-35s are represented as 60 isotropic point scattering centers and the F-14s as 50 such centers. Simulation data were obtained by assuming that two Su-35s and one F-14 were flying in a formation in the [−1 −1 0] direction with velocity v = 300 m/s and acceleration a = 10 m/s2 starting from the initial positions [ x1 y1 z1 ] = [ 0.310 50.0 3.0 ] km (first Su-35), [ x2 y2 z2 ] = [ 0.335 50.0 3.0 ] km (second Su-35), [ x3 y3 z3 ] = [ 0.285 50.0 3.0 ] km (F-14). Reflected signals were collected using (1) and (3). The simulation used pulse repetition frequency (PRF) = 2 kHz, center frequency f0 = 9.15 GHz, B = 200 MHz (0.75-m range resolution), and sampling rate = 512 MHz.
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
PARK et al.: SEGMENTATION OF ISAR IMAGES OF TARGETS MOVING IN FORMATION
2105
Fig. 6. Targets composed of scattering centers. (a) Three-dimensional representation of Su-35 consisting 60 point scattering centers. (b) Three-dimensional representation of F-14 consisting 50 point scattering centers.
Pulses were transmitted for 4.40 s (9800 pulses) to obtain a cross-range resolution approximate to the down-range resolution; to speed computation, they were downsampled to 128 equally spaced pulses. Then, the range profile history obtained from the given radar and motion parameters and the image used for the cost function were constructed using the proposed method (Fig. 7). The signal-to-noise ratio used was assumed to be 10 dB, and η to calculate the cost function was set to 50. The parameters used in PSO are population size = 50, number of generations = 20, φ = 0.5, and c1 = c2 = 1.49. The order of the polynomial in (5) was set to three; to find the solution more effectively, we set N/(2M ) ≤ a1 ≤ N/M , −1/(16M ) ≤ a2 ≤ 1/(16M ), and −1/(16M 2 ) ≤ a3 ≤ 1/(16M 2 ), where M is the number of pulses (128) and N is the number of range bins. In the first range-Doppler algorithm, the segmented alignment (Fig. 3) approach was used because the target trajectory is long and a single equation (5) to fit the trajectory cannot be easily derived. For comparison, alignment was performed using the proposed segmented alignment method and the conventional entropy method. In the proposed method, we used
Fig. 7. Range profile history and image for the cost function. (a) Range profile history. (b) Image used for the cost function.
six segments; the polynomial order for each segment was set to three. ISAR images were derived using each method after phase adjustment. Running the program written in Matlab R2007a in Windows XP on an Intel Quadcore processor, the computation time for the proposed alignment was 6.7712 s due to the need to calculate and align several segments, but the segmented alignment successfully aligned the range profiles [Fig. 8(a)] and yielded an ISAR image showing the three targets [Fig. 8(c)]. This result demonstrates that a trajectory that cannot be represented by one polynomial can be modeled by dividing it into several piecewise segments. In contrast, when using the entropy cost function, range bins with high amplitudes in the middle region were aligned [Fig. 8(b)] as that in Fig. 1. As a result, this method yielded an ISAR image [Fig. 8(d)] that shows only some high-valued pixels and does not yield easily identifiable images. Finally, each target was segmented using a window of 35 × 70 pixels in the second step and then enhanced in the third
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
2106
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 8. Comparison of the alignment results (segmented alignment) and ISAR images. (a) Segmented alignment. (b) Range alignment using the entropy cost function. (c) ISAR image computed using (a). (d) ISAR image computed using (b).
step (Figs. 9 and 10). In the third step, the range profiles were aligned using the entropy cost function because each segmented image contains only one target. This step improved the alignment of the range profile history (Fig. 9). Each target (Fig. 10) shows better focus than its original image [Fig. 8(c)]. IV. C ONCLUSION In this paper, we have proposed a new method to segment ISAR images of targets flying in a formation. It is composed of three steps: 1) an initial range-Doppler imaging of the whole set of targets; 2) segmentation of this initial image into component images; and 3) a second range-Doppler imaging, in which each ISAR image is enhanced using range alignment and phase adjustment. For the first range-Doppler algorithm, which is the main subject of this paper, we showed that the existing cost functions fail to provide correct range alignment for multiple targets
closely spaced flying in a formation because, in these methods, range bins with high amplitudes contribute more to the cost function than do bins with low amplitudes. We proposed a new range alignment method that solves this problem by constructing a new image whose values at scattering center locations are all adjusted to one, giving the same weight on the alignment to each scattering center, and by representing range shifts using a polynomial function. Coefficients of this function were estimated using PSO and a cost function which is the sum of the amplitudes of the adjusted pixels on the new image that occur along the curve described by the polynomial in the target region. Segmented alignment was proposed to cope with any mismatch between the trajectory and the polynomial. For the second step, a window with the same size as the target was used to partition each segment into subsegments. In this step, a binary image was constructed by converting pixels that are higher than the average power to one and the others to zero. Then, the first element with a value of one was located and
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
PARK et al.: SEGMENTATION OF ISAR IMAGES OF TARGETS MOVING IN FORMATION
2107
Fig. 9. Range profile histories (F-14) before and after alignment using the entropy cost function. (a) Before alignment. (b) After alignment.
used as the center of the window. COM was calculated using the pixels within the window; then, a new window centered on the COM was used to calculate a new COM. This procedure was repeated until the entire target was contained within the window. The third step was used to enhance the ISAR images. Simulation results derived from targets flying in a formation at a long range demonstrate that the proposed alignment method is superior to the entropy minimization method and that the proposed segmentation method can successfully image each target. The final goal of our study for generating ISAR images of multiple targets in a formation and their segmentation focuses on the noncooperative target recognition research. In a real situation, aircraft targets flying in a formation are usually detected in a single radar beam to hide their number, but associated articles have rarely found in open literatures about ISAR imaging. To identify target types in that situation, a robust ISAR imaging and segmentation algorithm is essential, and this paper addresses that problem. Currently, an experiment
Fig. 10. Segmented and enhanced ISAR images. (a) ISAR image of F-14 [bottom in Fig. 8(c)]. (b) ISAR image of Su-35 [middle in Fig. 8(c)]. (c) ISAR image of Su-35 [top in Fig. 8(c)].
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.
2108
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
to generate ISAR imaging of a real flying target has been successfully completed in Korea, and its extension to multiple aircraft targets will be conducted in recent future. R EFERENCES [1] C. C. Chen and H. C. Andrews, “Target-motion-induced radar imaging,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-16, no. 1, pp. 2–14, Jan. 1980. [2] M. M. Menon, E. R. Boudreau, and P. J. Kolodzy, “An automatic ship classification system for ISAR imagery,” MIT Lincoln Lab. J., vol. 6, no. 2, pp. 289–308, 1993. [3] D. A. Ausherman, A. Kozma, J. L. Walker, H. M. Jones, and E. C. Poggio, “Development in radar imaging,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-20, no. 4, pp. 363–400, Jul. 1984. [4] A. Wang, Y. Mao, and C. Chen, “Imaging of multi-targets with ISAR based on the time–frequency distribution,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 1986, vol. 5, pp. 173–176. [5] X. Wu and Z. Zhu, “Simultaneous imaging of multiple targets in an inverse synthetic aperture radar,” in Proc. IEEE Nat. Aerosp. Electron. Conf., May 1990, vol. 1, pp. 210–214. [6] Y. Mao, G. Chen, and J. Wang, “SAR/ISAR imaging of multiple moving targets based on combination of WVD and HT,” in Proc. CIE Int. Conf. Radar, Oct. 8–10, 1996, pp. 342–345. [7] Y. Kazuhiko, I. Masafumi, F. Takahiko, and K. Tetsuo, “An ISAR imaging algorithm for multiple targets of different radial velocity,” Electron. Commun. Jpn., Part I: Commun., vol. 86, no. 7, pp. 1–10, Mar. 2003. [8] X. Li, G. Liu, and J. Ni, “Autofocusing of ISAR images based on entropy minimization,” IEEE Trans. Aerosp. Electron. Syst., vol. 35, no. 4, pp. 1240–1251, Oct. 1999. [9] J. M. Munoz-Ferreras, F. Perez-Martinez, J. Calvo-Gallego, A. Asensio-Lopez, B. P. Dorta-Naranjo, and A. Blanco-del-Campo, “Traffic surveillance system based on a high-resolution radar,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 6, pp. 1624–1633, Jun. 2008. [10] D. Zhu, L. Wang, Y. Yu, Q. Tao, and Z. Zhu, “Robust ISAR range alignment via minimizing the entropy of the average range profile,” IEEE Geosci. Remote Sens. Lett., vol. 6, no. 2, pp. 204–208, Apr. 2009. [11] F. Berizzi and G. Cosini, “Autofocusing of inverse synthetic radar images using contrast optimization,” IEEE Trans. Aerosp. Electron. Syst., vol. 32, no. 3, pp. 1191–1197, Jul. 1996. [12] J. Wang, X. Liu, and Z. Zhou, “Minimum-entropy phase adjustment for ISAR,” Proc. Inst. Elect. Eng.—Radar, Sonar Navig., vol. 151, no. 4, pp. 203–209, Aug. 2004. [13] R. Hassan, B. Cohanim, and O. Weck, “A comparison of particle swarm optimization and the genetic algorithm,” in Proc. 46th AIAA/ASME/ ASCE/AHS/ASC Struct., Struct. Dyn. Mater. Conf., Apr. 2005, pp. 18–21. [14] H. Lin, J. Si, and G. P. Abousleman, “Knowledge-based hierarchical region-of-interest detection,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2002, vol. 4, pp. IV-3628–IV-3631. [15] H. Lin and G. P. Abousleman, “Hierarchical region-of-interest detection,” Opt. Eng., vol. 45, no. 7, pp. 077 201-1–077 201-11, Jul. 2006. [16] O. Lankoande, M. M. Hayat, and B. Santhanam, “Segmentation of SAR images based on Markov random field model,” in Proc. IEEE Int. Conf. Syst., Man, Cybern., Oct. 2005, vol. 3, pp. 2956–2961.
Sang-Hong Park received the B.S. and M.S. degrees in electronic engineering from Pohang University of Science and Technology (POSTECH), Pohang, Korea, in 2004 and 2007, respectively, where he is currently working toward the Ph.D. degree in electrical engineering. His research interests are in the areas of radar target imaging and recognition, radar signal processing, target motion compensation, pattern recognition using artificial intelligence, and RCS prediction.
Hyo-Tae Kim received the B.S. and M.S. degrees in electronics engineering from Seoul National University, Seoul, Korea, in 1978 and 1982, respectively, and the Ph.D. degree in electrical engineering from The Ohio State University, Columbus, in 1986. After his graduate work with the ElectroScience Laboratory, The Ohio State University, he joined the faculty of Pohang University of Science and Technology (POSTECH), Pohang, Korea, where he is currently a Professor. His research activities and interests are in the areas of antennas, EM scattering, EMI/EMC, and radar target identification.
Kyung-Tae Kim received the B.S., M.S., and Ph.D. degrees in electronic engineering from the Department of Electronic and Electrical Engineering, Pohang University of Science and Technology (POSTECH), Pohang, Korea, in 1994, 1996, and 1999, respectively. In 2002, he joined the faculty of the Department of Electronic Engineering, Yeungnam University, Gyeongsan, Korea, where he is currently an Associate Professor. He has published more than 80 papers in refereed journals and conference proceedings related to radar signal processing. He has many experiences on various signal processing areas such as spectral analysis, time–frequency transform, pattern recognition, and soft computing with applications to radar signal processing. His current research area of interests includes SAR/ISAR imaging with motion compensation, NCTR using HRR profiles and ISAR imaging, and SAR ATR, and related research projects funded by government and defense industries are in progress.
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on May 18,2010 at 03:59:56 UTC from IEEE Xplore. Restrictions apply.