An Efficient Multi-Target SAR ATR Algorithm

An Efficient Multi-Target SAR ATR Algorithm L.M. Novak, G.J. Owirka, and W.S. Brower MIT Lincoln Laboratory Abstract MIT Lincoln Laboratory has developed the ATR (automatic target recognition) system for the DARPA-sponsored SAIP program; the baseline ATR system recognizes 10 GOB (ground order of battle) targets; the enhanced version of SAIP requires the ATR system to recognize 20 GOB targets. This paper compares ATR performance results for 10- and 20-target MSE (meansquared error) classifiers using medium-resolution SAR (synthetic aperture radar) imagery. Introduction MIT Lincoln Laboratory is responsible for developing the ATR system for the DARPA-sponsored SAIP program. SAIP supports new sensor platforms such as the GLOBAL HAWK system, which gathers wide area SAR stripmap imagery at medium resolution (1.0 m 1.0 m) and SAR spotlight imagery at high resolution (0.3 m 0.3 m). The classification stage of the SAIP ATR provides target recognition at both medium and high resolution. In high-resolution spotlight mode, conventional 2D- FFT image formation processing is used to construct the 0.3 m 0.3 m resolution SAR imagery that is used to perform target recognition. In medium-resolution stripmap mode, a superresolution image formation algorithm is used to enhance SAR image resolution prior to performing target recognition. This image formation algorithm enhances the resolution of the 1.0 m 1.0 m imagery to approximately 0.5 m 0.5 m. Reference [1] compared ATR performance results for the 10- and 20-target MSE classifiers using high-resolution (0.3 m 0.3 m) SAR imagery to perform target recognition. This paper focuses on a comparison of ATR performance for 10- and 20-target MSE classifiers using medium- resolution (1.0 m 1.0 m) SAR imagery to perform target recognition. The Lincoln Laboratory baseline ATD/R system, which is depicted in Figure 1, consists of three basic data-processing stages: (1) detection, (2) discrimination, and (3) classification. In the detection and discrimination stages, the goal is to eliminate from further consideration any portions of the input imagery not containing targets, but simultaneously to allow all portions of the imagery containing targets to pass through to the classification stage. Classification is performed on each input subimage by finding the best matching target image from a database of stored target reference images or templates. The purpose of classification is to categorize the object in the input image either as a target of interest (of which there may be many types, e.g., T72 tank, M109 howitzer, etc.), or as an uninteresting clutter object. Objects that are determined to be in the former category are labeled with the target type corresponding to the best matching template, whereas objects in the latter category are simply labeled unknown. When the basic classification algorithm is performed, the best matching template is declared to be the one that yields the smallest MSE value with respect to the input image. However, as implied in Figure 1, this template-matching algorithm is actually performed twice within the new, efficient multiresolution classification stage. The initial MSE preclassifier is implemented using imagery that has the inherent sensor resolution (assumed in this paper to be 1.0 m 1.0 m). This preclassifier provides coarse classification information that is used to reduce the final, higher resolution template search space (target type, aspect angle, and spatial offset). After preclassification, a superresolution technique [2] known as high-definition imaging (HDI) is applied to the input image before the image is passed to the final high-resolution MSE classification algorithm. This multiresolution architecture for the classification subsystem reduces computational expense. Background and Data Description The synthetic aperture radar imagery used in these studies was provided to Lincoln Laboratory by Wright Laboratories, WPAFB, Dayton, Ohio. The data were gathered by the Sandia X- band, HH-polarization SAR sensor at two different sites in support of the DARPA-sponsored MSTAR program [3]. The first MSTAR collection (MSTAR #1) took place in fall 1995 at Redstone Arsenal, Huntsville, Alabama; the second MSTAR collection (MSTAR #2) took place in fall 1996 at Eglin AFB, Ft. Walton Beach, Florida. In each collection, a large number of military targets were imaged in spotlight mode, over 360 of target aspect, and at 0.3 m 0.3 m resolution. For the studies presented in this paper, these SAR data were processed to have 1.0 m 1.0 m resolution. Our initial studies [4] evaluated the performance and summarized the results of a 10-target MSE classifier using imagery of the 18 distinct targets contained in the MSTAR #1 data set. Figure 2 shows a typical SAR spotlight image (processed to 1.0 m 1.0 m resolution) of the Redstone Arsenal target array. We used 15 depression target images to construct a 10-target classifier. The classifier was trained by constructing classifier templates using SAR images of the following targets: BMP2#1, M2#1, T72#1, BTR60, BTR70, M1, M109, M110, M113, and M548 (see Figure 3). The target array shown in Figure 2 includes three versions each of the BMP2 armored personnel carrier, the M2 infantry fighting vehicle, and the T72 main battle tank. The three T72 tanks varied significantly from tank to tank; T72#1 used in training the classifier had skirts along both sides of the target; T72#2 had fuel drums (barrels) mounted on the rear of the tank; T72#3 had neither skirts nor barrels. The classifier was tested using the remaining 8 targets that were not used in training the classifier: two BMP2s, two M2s, two T72s, the HMMWV, and the M35. In our initial studies, the HMMWV and the M35 were used as confuser vehicles (i.e., vehicles not included in the set of 10 targets that the classifier was trained to recognize); the other 6

test targets provided independent classifier testing data (data not used in classifier training). One important conclusion gleaned from these initial studies [4] was that the ability to correctly classify the independent T72 targets depended strongly on how closely the target configuration matched that of the tank used in training the classifier. Because of the presence of the fuel drums located on the rear of the tank, T72#2 was called unknown a significant number of times. Using additional T72 tank data from MSTAR #2, we demonstrated that intraclass variability is a very important issue for classifier design [1]. This paper compares the performance of the 10-target MSE classifier with the performance obtained from the 20-target MSE classifier using medium-resolution (1.0 m 1.0 m) SAR imagery. To implement the 20-target classifier we combined 11 target types imaged during the MSTAR #1 collection with 9 target types imaged during the MSTAR #2 collection (both data sets at 15 depression). Figure 4 shows a typical SAR spotlight image (processed to 1.0 m 1.0 m resolution) of the (MSTAR #2) Eglin AFB target array. Photographs of the targets used to implement the 20-target classifier are shown in Figure 5. The 10- and 20-target classifiers were implemented by constructing 72 templates per target. These templates were obtained by using the target images which were gathered every degree in aspect around the target. Five consecutive images were then averaged to form 72 5 -average images per target. The templates were then obtained by isolating the clutter-free target pixels from each 5 -average image, providing 72 5 -templates spanning a total 360 aspect coverage per target. Efficient Classifier Implementation We developed a computationally efficient implementation of the MSE classifier for the SAIP system to provide significantly increased speed in the ATR function with only a marginal loss in ATR performance. As shown in Figure 1, the high-resolution MSE classifier is preceded by a preclassifier stage that performs a coarse MSE classification on 1.0 m 1.0 m resolution data. This reduced resolution MSE preclassifier provides an estimate of the pose (aspect angle of the target) and an estimate of the target s true class. This information is passed to the more computationally intensive high-resolution MSE classifier and is used to limit the search space over target aspect and target type, which results in a more computationally efficient ATR algorithm [4]. Figure 6 shows a cumulative error probability curve of the 20-target MSE preclassifier pose-estimation error in degrees for 1.0 m 1.0 m resolution target data. Because each template represents 5 of aspect angle, a pose error of 20 in angle corresponds to ±4 templates. The curve in Figure 6 indicates that approximately 95% of the time the correct pose is contained in the ±4-template search space. Note that a 180 ambiguity is included in the pose estimates because these targets are nearly symmetric when facing forward or backward. Therefore, a ±4-template pose estimate with the 180 ambiguity yields a total of 18 templates to be searched at higher resolution. Thus the higher-resolution MSE classifier does not have to search all 72 templates per target; rather, it searches a much smaller subset of the high-resolution template set. Figure 7 presents a plot of the probability that the correct target class is contained in the top N MSE scores for 1.0 m 1.0 m resolution imagery. For this study, the top score gave the correct class only 31.7% of the time. The correct class for the 20-target classifier was contained in the top 10 scores approximately 94.2% of the time. The curves in Figures 6 and 7 show that the 20 pose error angle and the top 10 scores from the preclassifier can be used to prune the high-resolution MSE classification search space with only a small degradation in performance. Performance Results This section of the paper summarizes the ATR performance achieved by the 10- and 20-target MSE classifiers using 1.0 m 1.0 m resolution SAR imagery. Both classifiers were initially tested using the 6 independent targets from the MSTAR #1 collection. The results of these evaluations are summarized in Table 1, which presents the classifier confusion matrices for the 10-target classifier trained using MSTAR #1 data and tested on the 6 MSTAR #1 independent test targets (top) and for the 20-target classifier tested on 6 MSTAR #1 independent test targets (bottom). When the 10-target classifier was tested using the independent MSTAR #1 test data, an average probability of correct classification of 66.2% was achieved against the 6 independent targets. Note, however, that the performance for the T72 tank with fuel drums on the rear (T72#2) was somewhat reduced; 123 images of the 195 total were correctly classified, while 44 images of the 195 total were declared unknown by the classifier. When the 20-target classifier was tested using the same independent MSTAR #1 test data, the average probability of correct classification degraded slightly to 60.7%. The number of T72#2 targets correctly classified by the 20-target classifier was only 104 of the 195 total, while 31 images were declared unknown. Both classifiers were then tested using independent test data (three BTR70s and four M109s) in controlled configurations from the MSTAR #2 collection. Table 2 presents the classifier confusion matrices for the original 10-target classifier tested on the 7 MSTAR #2 independent test targets (top) and for the 20-target classifier tested on the 7 MSTAR #2 independent test targets (bottom). The probabilities of correct classification against these independent test data are 77.3% and 70.3% for the 10- and 20- target classifiers, respectively. This test illustrates that classifier templates developed from the MSTAR #1 collection work equally well when tested against these independent test target images from the MSTAR #2 collection. The MSTAR #2 collection imaged eight T72 tanks in a variety of configurations, as described in Table 3. We tested the 10- and 20-target classifiers using target images of seven of the independent T72 tanks from the MSTAR #2 collection. Note that a single T72 tank from the MSTAR #1 collection was used to train both classifiers; its configuration was skirts/no barrels (S/NB; i.e., skirts along both sides of the tank but no fuel drums mounted at the rear). When both classifiers were tested against seven of the independent T72 tanks from the MSTAR #2 collection,

significantly degraded classifier performance was observed. As shown in Table 4, the probabilities of correct classification against these test data are 52.3% and 36.4% for the 10- and 20-target classifiers, respectively. As shown in Table 4, the 10-target classifier rejected a large number of T72 tank images (424 of the total 1918), declaring them unknown. The confusion matrix indicates that 93 images of T72 #7 were rejected, 93 images of T72 #6 were rejected, and 83 images of T72 #5 were rejected. Note that T72 #5, #6, and #7 were all configured with fuel barrels and T72 #7 was also configured with reactive armor. The 20-target classifier confusion matrix presented in Table 4 illustrates that only 698 T72 test images were correctly classified of the total 1918; also, 263 test images were declared unknown. Increasing the number of target classes from 10 to 20 resulted in many more T72 test images being incorrectly classified. Table 4 (bottom) indicates that for the 20-target classifier the M60 and T62 tank classes were a considerable source of confusion for the T72 test inputs. For T72 #5 and #6 alone, 149 T72 test inputs were incorrectly classified as T62 tanks. Because the T72 was trained using only the S/NB configuration target from the MSTAR #1 collection, it was decided that the various T72 configurations should be investigated more carefully. SAR images and optical photographs of the MSTAR #1 and MSTAR #2 targets were used to investigate target configuration, especially for the many T72 variants and the T62. We compared T72 test images from MSTAR #2 with the T72 training images from MSTAR #1 and the T62 target images. Many test images of the T72 tanks that were configured with fuel barrels, such as MSTAR #2 T72 #5, #6, and #7 had scatterering signatures that were more similar to the T62 templates than the T72 templates. This discrepancy was explained by observing that the images used to construct the T62 templates were configured with skirts/barrels and the images used to construct the T72 templates were configured with skirts/no barrels. These observations prompted an experimental modification of the classifiers. We speculated that augmenting the classifier template sets with an additional template set of a T72 tank having a no skirts/barrels configuration would improve overall classification performance. The study using the MSTAR #2 T72 tanks (summarized in Table 4) was repeated for the 10- and 20-target classifiers with the additional (NS/B) T72 templates. For clarity, we will refer to these classifiers as the 11- and 21-target classifiers even though they still only identify 10 and 20 unique target types. The 10-target classifiers discussed earlier (Tables 1, 2, 4) used 720 preclassifier templates and 720 high-resolution templates; the 20-target classifiers used 1440 preclassifier templates and 1440 highresolution templates; the 11- and 21-target classifiers use an additional 144 templates for the T72 variant. Table 5 summarizes the results of the 11- and 21-target classifiers; the probabilities of correct classification improved to 75.6% and 63.9% for the 11- and the 21-target classifiers, respectively. A full evaluation of the performance of the 11- and 21-target classifier implementations was performed by combining the independent test inputs from the MSTAR #1 and MSTAR #2 data sets. Table 6 presents confusion matrices for 5195 independent test inputs. The leftmost column denotes the target type, followed by the number of different-serial-numbered targets of each type used in the performance evaluation. For example, there were 9 different-serial-numbered T72s included in this final performance summary. Since the total number of test inputs varies with each target type, the confusion matrix entries have been converted to percentages. As Table 6 shows, the average probabilities of correct classification are 74.4% and 66.2% for the 11- and 21-target classifiers, respectively. The results of this evaluation indicate that classification performance can be maintained even with significant target configuration variability if additional templates with the appropriate configuration are incorporated into the classifier. Of course, the addition of classification templates to compensate for a target configuration variation does increase the overall storage requirement and computation of the classifier. Summary This paper compared the performance of 10- and 20-target, template-based, MSE classifiers. Both classifiers were developed at Lincoln Laboratory in support of the SAIP program. The classifiers use medium-resolution (1.0 m 1.0 m) data processed using a new superresolution imaging technique high-definition imaging in an efficient multiresolution architecture. Highdefinition imaging improves overall classification performance while the multiresolution implementation reduces the computational load. System performance was evaluated using a significant number of tactical military targets (5195 test images). The results of these evaluations show that the number of target classes can be increased from 10-target classes to 20-target classes with only a small decrease in target recognition performance. The correct classification performance for the final 10- and 20-target classifiers was 77.4% and 66.2%, respectively. The results of these evaluations also show that significant target configuration variability can decrease interclass separability and degrade performance; however, additional reference templates can be used to mitigate these effects. References 1. L.M. Novak, et al, Performance of a 20-Target MSE Classifier, SPIE Conference, Orlando, Fla., April 1998. 2. G.R. Benitz, High-Definition Vector Imaging for Synthetic Aperture Radar, Asilomar Conf., Pacific Grove, Calif., November 1997. 3. MSTAR Program Technology Review, Denver, Colo., November 1996. 4. L.M. Novak, et al, The ATR System in SAIP, Lincoln Laboratory Journal, Vol. 10, No. 2, 1997.

Figure 1. Block diagram of the Lincoln Laboratory baseline ATR system. Figure 4. Typical SAR spotlight image (1.0 m 1.0 m resolution) of the Eglin AFB target array. Figure 2. Typical SAR spotlight image (1.0 m 1.0 m resolution) of the Redstone Arsenal target array. 304110-22P BMP2 #1 M2 #1 T72 #1 BTR60 BTR70 M548 BMP2 #2 M2 #2 T72 #2 M1 M109 M110 BMP2 #3 M2 #3 T72 #3 M113 M35 HMMWV Figure 5. The targets from the MSTAR #1 and #2 collections used to train the 20-target classifier. Table 3 Intraclass variability matrix (seven T72 tanks from the MSTAR #2 data set) 303160-1B T72 Intra-class Variability Matrix Figure 3. Photographs of the 18 targets from the MSTAR #1 collection. The classifier was trained with 10 targets (BMP2#1, M2#1, T72#1, BTR60, BTR70, M548, M1, M109, M110, M113). Six independent targets (BMP2#2, M2#2, T72#2, BMP2#3, M2#3, T72#3) and two confuser targets (M35, HMMWV) provided test data for the classifier. Notation Configuration of Target S/B Skirts/barrels (fuel drums) S/NB Skirts/no barrels NS/B No skirts/barrels NS/NB No skirts/no barrels S/B/A Skirts/barrels/reactive armor

Figure 6. Cumulative error probability versus pose error with 1.0 m 1.0 m resolution target data. Figure 7. The probability that the correct target class is contained in the top N MSE scores with 1.0 m 1.0 m resolution target data. Table 1 Confusion matrices for the 10-and 20-target classifiers using 1.0 m 1.0 m resolution imagery (test inputs are six independent targets from the MSTAR #1 data set)

Table 2 Confusion matrices for the 10- and 20-target classifiers (test inputs are seven independent targets from the MSTAR #2 data set) Table 4 Confusion matrices for the 10- and 20-target classifiers (test inputs are seven T72 tanks from the MSTAR #2 data set)

Table 5 Confusion matrices for the 11- and 21-target classifiers (test inputs are seven T72 tanks from the MSTAR #2 data sets) Table 6 Confusion matrices for the 11- and 21-target classifiers (test inputs are a composite of MSTAR #1 and MSTAR #2 data sets)