Multisensor Data Fusion for Surface Land-Mine Detection

Transcription

1 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY [10] T. Takagi and M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Trans. Syst., Man, Cybern., vol. SMC-15, Jan [11] J. Valente de Oliveira and J. M. Lemos, Long-range predictive adaptive fuzzy relational control, Fuzzy Sets Syst., no. 70, pp , [12] J. Valente de Oliveira, A design methodology for fuzzy system interfaces, IEEE Trans. Fuzzy Syst., vol. 3, no. 4, pp , [13] J. Valente de Oliveira and J. M. Lemos, Speeding up fuzzy relational identification: The RLS approach, in Proc. 6th Int. Fuzzy Systems Assoc. World Cong., São Paulo, Brazil, 1995, pp [14] J. Valente de Oliveira, Prediction using relational systems, in Fuzzy Modeling: Paradigms and Practice, W. Pedrycz, Ed. Norwell, MA: Kluwer, 1996, pp Multisensor Data Fusion for Surface Land-Mine Detection Arthur Filippidis, L. C. Jain, and N. Martin Abstract Receiver-operating curves have been used to examine a novel target-recognition system using a number of knowledge-based techniques to automatically detect surface land mines present in 30 sets of thermal and multispectral images. A summary of results, graphed at a probability of detection greater than or equal to 96%, shows the false-alarm rates (FAR s) obtained using various combinations of fusing sensors and neural classifiers averaged over the 30 images. Results show that using two neural-network classifiers on the input textural and spectral characteristics of selected multispectral bands, we obtain FAR s of approximately 3%. Using polarization-resolved images only, we obtain FAR s of 1.15%. Fusing the best classifier output with the polarization-resolved images, we obtain FAR s as low as 0.023%. This result has shown the large improvement gained in the fusion of sensors. Also, an improvement is shown by comparing these results with those reported in an existing commercial system published in an internal report. Index Terms Fusion, receiver-operating curves. I. INTRODUCTION Land mines pose problems for the military (by restricting their mobility) and for civilian populations, who are at risk long after conflicts have passed. Many of the land-mine detection systems under development employ two or more types of sensors with their outputs fused together to maximize the detection performance while minimizing false alarms. Sensors are often vehicle-mounted or even operated from low-flying helicopters. The primary requirement of any land-mine detection system is a high probability of detection (Pd) and a low false-alarm rate (FAR). One of the aims of this study is to provide a quantitative demonstration of the benefits of using multiple sensors for the detection of surface land mines. In particular, it will show a reduction in the FAR through the fusion of two sets of imagery from an infrared (IR) sensor with a rotating polarizer attached and a digital multispectral camera. Some very recent work in surface mine detection using a polarimetric IR sensor is described in [1]. In this paper, the authors use a polarimetric IR camera for detecting man-made objects (such as mines) by detecting the regularly polarized man-made objects, which contrast with background clutter. The obvious problem is that it will not distinguish land mines from other man-made clutter. In the novel automatic target recognition (ATR) system described in Section III (shown in Fig. 2), preprocessing the polarized image using a background-discrimination algorithm [2] before using knowledge-based techniques for detection, has provided a novel way of automating the detection of targets in polarization images. The ATR system also describes a new way of distinguishing the land mine over other man-made clutter. One of the successful commercial mine-detection systems is marketed by Marietta Electronics and Missiles [3]. The system uses a forward-looking, 8 12 m infrared sensor (FLIR) that looks for target characteristics in terms of area, perimeter, moments, and intensity measures. The ATR system is based on the classical ATR approach detailed in Fig. 1. They use a combination of three neural-network approaches, supervised real-time learning networks, and unsupervised real-time learning to cover a number of different scenarios of known and unknown mines, clutter, and terrain. The preprocessing stage of the sensor data used to extract the features is the crucial stage. It consisted of local filtering, histogram equalization, linear expansion, contrast stretching, feature extraction, image enhancement, segmentation, and a prescreener. It can be employed at up to 80 ft stand-off range. The image-enhancer filters perform mainly low-frequency filtering and enhance the quality of the image. The segmenter extracts edges from the scene using a Sobel-edge operator followed by a weighted Generalized Hough Transform to perform matched filtering. This is followed by feature extraction, obtaining size, shape, and thermal contrast with an algorithm (not given) developed by Martin Marietta. The goal of this work is to investigate the detection of surface land mines given multiple registered images of the mined area, obtained from a suite of visible to IR wavelength sensors. We will be looking into the automatic detection of surface land mines. The novel approach takes the outputs from two different imaging sensors: a multispectral (optical) camera and a thermal IR imager fitted with a rotating polarizer. The target information from the two images is fused together using a fuzzy rule-based system [6]. Compared to the earlier system discussed in [3], the new approach is more suited to above-surface land mines, because the multispectral sensor will identify only surface targets, and the rotating polarization filter attached to the thermal imager lens will now highlight surface targets by its particular polarization signature. The technique reported here used the novel architecture shown in Fig. 2, which combines sensors and signal-processing algorithms using neural networks and a fuzzy rule-based fusion algorithm to detect all the land mines and reduce the FAR. A comparison of the best results obtained in experiments using both ATR techniques is shown in Table I of Section V. Table I shows a large improvement in the reduction of the FAR using the fusion system. II. OBJECTIVES Manuscript received September 29, 1998; revised October 26, 1998 and March 23, A. Filippidis and N. Martin are with the Land Operations Division, Defence Science Technology Organization, Salisbury, Australia ( arthur.filippidis@dsto.defence.gov.au). L. C. Jain is with the Knowledge Based Intelligent Engineering Systems Centre, University of South Australia, Adelaide, Australia ( L.jain@unisa.edu.au). Publisher Item Identifier S (00) The goal of this work is to investigate the detection of surface land mines given multiple registered images of the mined area obtained from a suite of visible to IR wavelength sensors. Between two and six of the above mines described in different combinations and locations were used together with man-made clutter. If an area has been subject to military action, then it may be littered with battle debris consisting of spent cartridge cases, metal fragments, and destroyed equipment [4]. These features create a high degree of clutter when viewed /00$ IEEE

2 146 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY 2000 Fig. 1. Classical ATR system used by Martin Marietta Inc., taken from [3]. using thermal imagery. In order to characterize man-made clutter, objects such as four square markers were used, as well as aluminum, teflon, and brass blocks. Natural clutter consists of seasonal grasses, twigs, and small shrubs on a hard-packed red loam, as well as a small gravel and light-colored, sandy-soil area. Image processing using a background-discrimination algorithm [2] has been used on the polarized images to help distinguish the man-made targets from the natural clutter. The sensor suite consists of a digital camera capable of acquiring a range of visible and near-ir bands and the Agema themovision 900 radiometric thermal-imaging system, fitted with a rotating polarizer on the 8 12-m scanner. Utilization of IR polarization using the rotating polarizer in passive thermal imaging to help detect the surface mines is a novel research area, and as yet is still in its infancy. The fusion of the polarization results, together with the multispectral results, is discussed in [6]. Fig. 2. The ATR system uses a fuzzy rule-based fusion technique to combine land-mine identity attributes from the outputs of the MLP, ART2, and preprocessed thermal-polarization image. TABLE I AVERAGE FAR OF THE ATR SYSTEM (MLP FUSED WITH POLAR) IN FIG. 2IS COMPARED WITH THE FAR FOR THE BEST (AT P = 97:4%) AND WORST CASE [3] III. AUTOMATIC DETECTION USING MULTISENSOR DATA FUSION The self-organizing network ART2 [5] has been modified in a novel way [6] to provide a fuzzy-output value, which indicated the degree of familiarity (called familiarity flag in [6]) of a new analog input pattern compared to previous patterns stored in the long-term memory weights of the network. Hence, modified ART2 output [6] indicates the likelihood of a surface land mine with a value ranging from zero (nonsurface land mine) to one (surface land mine). The ATR system in Fig. 2 shows the outputs of the multilayer perceptron and the modified output of ART2 [6] to provide an analog value to a fuzzy rule-based fusion technique [6], which also uses a processed polarization-resolved image as its third input. In real time, these two classifier outputs indicate the likelihood of a surface land-mine target when presented with a number of multispectral and textural bands. The ATR system in Fig. 2 uses fuzzy rule-based fusion to combine complementary information derived from both sensors to produce an output image showing the likelihood of mine locations. The inputs to the fusion process are the classification outputs of modified ART2 and the MLP, together with the output of the processed IR polarization image. ART2 was selected for its robust ability to train targets in real time (in less than a minute in our case) and automate the system (unlike the MLP, which requires hours of training). Using the accuracy of the test data and the training of the mine data and the nonmine data, a genetic algorithm (GA) tool (described in Section IV) is used to find the optimum structure and inputs of the MLP neural network. The MLP and ART2 s inputs (using a 5 5 pixel window moving simultaneously across each of the spectral/textural bands shown in Fig. 2) are the average spectral characteristics from the 450-nm, 500-nm, 600-nm red and green bands, together with the three texture measures [7] (contrast, second angular momentum, and correlation) derived from the 450-nm band. The eight inputs to the neural networks are shown to the left of Fig. 2. Hence, identity attributes in the fusion algorithm derived from these two neural classifiers A and B (shown in Fig. 2) range from zero (no land mine) to one (land mine) and indicate the likelihood of a land mine at each pixel as the 5 25 window moves across the test image. Two polarization-resolved images are subtracted at two different polarization angles (0 and 90 ) to obtain the polarization image. It then is processed using a background-discrimination algorithm [2] to identify the man-made targets in the image. The preprocessing continues using a background-discrimination algorithm [2] at a number of threshold values (discussed later) to produce the polarization-processed image that will be used to obtain the area-identity attribute used in the fusion algorithm [6] (refer to Fig. 2). An 8 28 pixel window simultaneously moves across the processed-polarization image at the same pixel position as the other eight (5 25 pixel) windows move across the eight spectral/texture bands, inputting data into the two classifiers. The 8 28 pixel window (appropriately sized for the land mines in the processed polarized image at the particular stand-off range) is used to calculate the area of the processed image only if there are at least four pixels at the center of the window. For example, we count the number of black pixels connected together in the horizontal and vertical directions in the 8 28 pixel window as it moves across the entire image one pixel at a time. We assume a priori knowledge of the approximate size and area of the land mines for the digital camera viewing angle and stand-off range. Note that all the images are registered to the multispectral image, which is an (300 rows 2700 column pixels) image obtained using a digital camera. Hence, the location of mines, vegetation, and other man-made objects are within 4 5 pixel accuracy.

3 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY IV. MULTISPECTRAL IMAGERY AND TEXTURE PROCESSING A passive multispectral scanner is used to detect spectral intensity differences between surface land mines and man-made and natural clutter. However, it is limited to daylight hours of operation. The reason for using multispectral analysis is that it has the potential to discriminate between surface land mines and all other surface clutter. At best, thermal imagery and thermal polarization-resolved images will only be able to detect most of the surface targets and clutter. A neural-network classifier has been trained to discriminate against clutter and automatically detect land mines. One of the main reasons for concentrating on the neural-network classifier (the MLP in particular) is that it has been shown [8] to perform as well as the classical Bayesian classifiers for classes of multispectral data that are normally distributed. Moreover, where nonnormally distributed multispectral data is concerned (as is the case when classes of different areas or targets are combined as one class, such as three different mines), the MLP performs better, with an increase in accuracy of at least 5%. The digital camera provided several bands in the near IR (450, 500, 550, 600, and 650 nm) and a color-composite image. Different combinations of bands using three-texture parameters computed from the co-occurrence matrices [7] (second-angular momentum and correlation and contrast of a 5 25 window), together with their spectral values, were explored using the MLP and ART2 neural networks. Texture is an important property in detecting targets in images. The texture properties and gray tone in an image are primarily independent properties [9]. In the fusion process, these texture and gray-tone properties of the land-mine targets are combined and fed into the fusion algorithm as identity attributes A and B (as shown in Fig. 2). For this reason, even if the land-mine target is camouflaged in gray tone, it is difficult to camouflage these properties and texture simultaneously [9]. There is a large array of texture measures in the literature [10]. The choice of texture features should be based on application and the nature of the images used. One of the drawbacks of texture feature is that it cannot be used for detecting small targets in general. The object should cover at least a 3 23 window. The GA was run using the NeuroGenetic Optimizer (NGO) software package on a 100 MHz Pentium. NGO is a practical tool designed to engineer neural networks naturally [11]. This system helps to easily and quickly discover the best combination of data elements and neural-network architectures to build effective neural-network applications. Hence, NGO is an automation tool that will offload hours of effort onto computers. The NGO uses genetic algorithms to perform a combinatorial search [11] across all provided input variables and neural-network configurations (within user-specified constraints), and then creates, trains, and tests these networks to determine their accuracy. The basic steps the NGO goes through are as follows: 1) opening the data file containing spectral and textural characteristics of the mine and nonmine data; 2) loading it into memory; 3) building and validating training and test-data sets; 4) creating a population of candidate input variables and neural structures; 5) building the neural networks and training and evaluating them; 6) selecting the top networks; 7) pairing up the genetic material representing the inputs and the neural structure of these networks; 8) exchanging genetic material between them; 9) throwing in a few mutations for a flavor of random search; 10) going back into the training/testing cycle again. This continues for a defined number of cycles (generations) for a defined period of time or until a neural-accuracy goal is reached. The first image containing all the land mines and clutter was used to produce a training template in which polygon areas were marked around the positions of the landmines, selected areas of clutter, ground cover, and vegetation. Using this template, a source-data file was created. The source-data file (derived from the first image containing all possible mines, vegetation, ground cover, and possible clutter) used for the GA consisted of 711 records (or the number rows of data) and 13 fields (columns). The 13 input fields consisted of the five texture measures [7] derived from the 450-nm band (second angular momentum, contrast and correlation, cluster shade, and cluster prominence), together with the following spectral bands: 450 nm, 500 nm, 550 nm, 600 nm, 650 nm, and color bands (red, green, and blue). The output field consisted of a binary one or zero, indicating mine or nonmine target pixels. The 450-nm band was used to derive the three texture measures, as initial experiments (trial and error) indicated it produced better accuracy results than any of the other bands. Every two records were split to create 365 training records and 355 testing records. Parameters used in this run were the generation run of 50 and population size of 30. The minimum number of network training passes for each network was 20. The cutoff for network training passes was 50, and the limit on hidden neurons was ten. Selection was performed on the top 50%, and refilling of the population was done by cloning the survivors. Mating was performed using the tail-swap technique, and mutations were performed using the random-exchange technique at a rate of 25%. The optimum network and inputs were found on generation 43 after a run time of 39 h. The accuracy of the training set used by the GA was 83% (i.e., 83% of training data was correctly classified) and 84% (i.e., 84% of test data was correctly classified) on the test set. The network was an MLP back-propagation neural network that employed eight inputs (the 450 nm, 500 nm, 600 nm, red, green, second angular momentum, contrast, and correlation-input fields) and one hidden layer with eight neurons. The hidden neurons used a linear-transfer function. The one output neuron used a sigmoid-transfer function. Adaptive-resonance architectures such as ART2 are neural networks that self-organize stable pattern-recognition codes in real time in response to a sequence of analog (gray-scale) input patterns [5]. ART2 encodes, in part, by changing the weights or long-term memory traces, of a bottom-up adaptive filter. This filter is contained in pathways leading from a feature-representation field (F1) to a category-representation field (F2) whose nodes undergo cooperative and competitive interactions [5]. ART2 has eight input neurons on the input F1 layer and 15 neurons on the F2 output layer. In the training phase, at a vigilance value of (adjusted or fine tuned on mine-training data), ART2 activated ten category neurons in F2 to classify the eight input data fields presented (just as in MLP) in the training file as all the possible surface land-mine features. The long-term memory weights of this network were then stored for future use for testing on the 30 images. Hence, the modified ART2 [6] will now provide a fuzzy output from its familiarity flag in the testing phase, which indicates the degree of match of the new inputs to the encoded weights stored in its long-term memory. V. TESTING RESULTS OF THE ATR SYSTEM This section describes the results of the ATR land-mine detection system discussed in previous sections. A traditional method of comparing detection algorithms in the ATR system (in Fig. 2) is the receiver operating characteristic (ROC) curve. Although nonstatistical, it shows the relationship between the Pd and FAR s over a range of threshold values. In the ATR system described in Fig. 2, we compare the fusion process with either one or both of the neural-classifier outputs together with the polarization output, using the background-discrimination algorithm [2] at six different threshold values selected at one standard de-

4 148 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY 2000 TABLE II FAR ESTIMATES FOR THE THREE ESTIMATED PROBABILITIES OF DETECTION S P AT THE 95% CONFIDENCE INTERVAL FOR THE POSSIBLE FUSION STATES USING POLARIZATION Fig. 3. Four ROC s showing processed polarization (at six threshold values using the background-discrimination algorithm [2]) on its own and in comparison to MLP fused with polarization, ART2 fused with polarization, and MLP fused with both ART2 and polarization. In addition, average-classifier MLP and ART2 P and FAR are shown. viation from the mean of the man-made targets (in the polarization-resolved image). We also provide an intuitive statistical assessment of the ATR system, analyzed using standard frequentist-confidence intervals. A target is defined as one or more pixels in the area of a land-mine position and a nontarget as any pixel not in a land-mine position. The FAR using the 30 images represents a fraction of target declarations that were nontargets in the entire image (300 rows by 700 columns). The 95% confidence interval discussed in Fig. 5 and Table II represents the range of error rates that will be observed 95% of the time for equivalently distributed data. The ATR system is calibrated by identifying a number of known areas on all the selected land mines and clutter pixels of the first image to be used as templates for classifying the whole image and subsequent images. The sets of reference data (training sets) taken from the known areas are used to generate parameters that characterize the classes around the four different land mines, ground surfaces, vegetation, and clutter (i.e., mean red-pixel value or particular texture measure in a 5 5 pixel window). Average output results (for the 30 images) were obtained at the six threshold values and displayed in Figs. 3 and 4. The advantage of ART2 (unlike the MLP) is that it will train in real time and hence, it is possible for the ATR system to be recalibrated if the vegetation or lighting conditions change in the course of a real-time application. The four ROC s in Fig. 3 show the average (over the 30 images) processed-polarization results taken at six threshold values (using the Fig. 4. Three ROC s for MLP fused with polarization, ART2 with polarization, and ART2 fused with polarization and MLP. background-discrimination algorithm [2]). It shows this on its own and compared to the MLP fused with polarization, ART2 fused with polarization, and MLP fused with ART2 and polarization also taken at the same six polarization-threshold values. In additon, average-classifier MLP and ART2 Pd and FAR results are shown taken over the 30 images. The three ROC s in Fig. 4 show an expanded FAR scale of only the fusion results (i.e., MLP fused with polarization, ART2 fused with polarization, and all three MLP, ART2, and polarization fused together). The comparative results in Table II show the 95% confidence intervals for the FAR (of polarization only, together with the fusion of polarization and MLP or ART or MLP and ART combined, and averaged over the 30 images) given an estimated rate of detection (95.5%, 96.5%, and 97.5%). The confidence intervals provide an intuitive measure of the relative performance of the different combinations in the ATR system. In order of accuracy, the bar graph in Fig. 5 graphically describes a clearly observed summary of the best FAR at the various input combinations used in the ATR system at the lower and upper 95% confidence interval, and averaged over the 30 images for Pd 96:5%. The 95% confidence interval can be described as the range of error rates that will be observed 95% of the time for equivalently distributed data. A distinct improvement in accuracy is observed using the three fusion cases over polarization and the two classifiers on their own. Using a supervised classifier, the internal report by M. Marietta [3] indicated that the best result was obtained at a Pd of 97.4% and a FAR of 2.5%, and that the worst result was obtained at a Pd of 97.1% and a FAR of 2.9%. Table I compares the average FAR using the fused output of the MLP and polarization image in the ATR system obtained (over

5 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY Fig. 5. Graphical summary of the best FAR (of the various input combinations used in the ATR system) at the lower and upper 95% confidence interval, and averaged over the 30 images for P 96:5%. the 30 images) for these two Pd values (obtained from Martin Marietta [3]). The MLP fused with polarization produced the most accurate results. Similar land mines were used in both results. VI. DISCUSSION OF RESULTS Although the experimental design in collecting the images was not specific about the meteorological conditions under which the experiment should be conducted, they did require these conditions to be consistent. For this reason, experimental results were taken on a clear day within a few hours of midday. The receiver-operating curves in Fig. 3 show a marked increase in accuracy when using any fusion combination of each or both neural classifier at the six threshold values derived from processing the polarization-resolved image using the background discrimination algorithm [2]. Fig. 3 shows that the best results in the polarization and fusion cases in terms of Pd and FAR occurs at approximately the fourth (from the top of the curve) threshold point. Also, the difference in magnitude of the FAR increases markedly between the three fusion combinations and polarization only, as we approach Pd =1. The average Pd versus FAR, using the two neural classifiers, also provided a good comparison among themselves, the fusion, and polarization results. As indicated in [12], which specifically compared ART2 and MLP in an image-target detection problem, it is not surprising that the MLP achieved higher accuracy rates in both Pd and FAR over ART2 (by 0.2 and 0.25%, respectively, as shown in Fig. 4). This is due to the fact that its strength lies in the area of training in real time, which is an important factor in the future application of the ATR system in situations where lighting, vegetation, and clutter may vary. Although the three fusion ROC s are not distinguishable in Fig. 3, the large difference between them and the nonfusion cases is well highlighted. Fig. 4 provides a closer look at the three fusion ROC s, showing the very good accuracy rates achieved using the MLP and polarization combination. In terms of accuracy, ART2 fused together with the MLP and polarization is next, and ART2 fused with polarization is last. The probable reason for the three-case fusion (MLP, ART2, and polarization) being the second in terms of accuracy is that the results from the slightly better classifier (MLP) balance or null the slightly poorer results obtained from ART2. The FAR differences between the three fusion ROC s is as little as 0.025% at the elbow portion of the curve, showing the optimum accuracy results. In the performing-confidence interval analysis, the ATR surface land-mine detection system is analyzed using standard frequentist-confidence intervals. The 95% confidence intervals shown in Table II is calculated for the FAR, given an estimated rate of detection common to all ROC s 95.5%, 96.5%, and 97.5%, providing an intuitive measure of the relative performance between using polarization only and using fusion of polarization with the neural classifiers. The table shows a consistently improving FAR accuracy rate for the three fusion combinations (especially for MLP fused with polarization), in which both the polarization and multispectral sensors were used (as opposed to using the polarization sensor on its own). Using MLP fused with polarization, MLP fused with ART2 and polarization, and ART2 fused with polarization, we obtained FAR improvements at Pd =96:5% of approximately 1.27%, %, and 1.065%, respectively, over using polarization only. In summarizing all the results graphically averaged over the 30 images for Pd 96:5%, the bar graph (in Fig. 5) shows the vast improvements in FAR at the lower and upper 95% confidence intervals between the fusion combinations and nonfusion cases of polarization and neural classifiers on their own. Comparing the most accurate result (MLP fused with polarization) with the nonfusion cases (polarization, MLP, and ART2), the bar graph (refer to Fig. 5) shows approximate FAR improvement of approximately 1.27%, 2.7%, and 3.1%, respectively. VII. CONCLUSION In the ATR system, we have used sensor fusion to overcome the limitations of the individual sensors, and we have used processing techniques for the surface-mine detection task. When one sensor or processing technique did not provide all the necessary information (as was very much the case in our system), a complementary sensor or processing technique (i.e., classifier output) provided additional information, which reduced our overall FAR. Our experiments have clearly shown that the accuracy rate is higher using both sensors. Polarization will only detect all man-made targets (including land mines), and will not discriminate between land mines and other man-made targets, whereas multispectral analysis has the potential to identify and distinguish land mines and clutter but is prone to a high FAR. Once calibrated, the ATR system combines both polarization and multispectral functions to drastically reduce the FAR. This was prevalent, as shown in the bar graph in Fig. 5. The use of the real-time classifier ART2 will help in the future concept-demonstrator phase of this work. The ATR system itself will need to be recalibrated (preferably in real time) as vegetation and lighting condition change markedly.

6 150 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART C: APPLICATIONS AND REVIEWS, VOL. 30, NO. 1, FEBRUARY 2000 In conclusion, we have demonstrated that the fusion of the outputs derived from these sensors has been able to drastically reduce the FAR obtained in both the multispectral and polarization-resolved images in this experiment. Prototyping Structural Description Using an Inductive Learning Program Adnan Amin ACKNOWLEDGMENT The authors wish to thank the Land Operations Division of Defence Science and Technology Organization, Adelaide, Australia, for the technical and equipment support in the set up of the minefield. They are also thankful for the comments and corrections from Principal Research Scientist B. Seymour and Research Leader I. Heron. REFERENCES [1] B. A. Babour, S. Kordella, M. J. Dorsett, and B. L. Kerstiens, Mine detection using a polarimetric IR sensor, presented at the Detection of Abandoned Land-Mine, 1996, IEE, Conf. Pub. Paper 431. [2] N. Stacy, R. Smith, and G. Nash, Automatic target recognition for the INGARRA airborne radar surveillance system, Defence Science Technology Organization, Salisbury, Australia, Microwave Radar Division Internal Rep., Aug [3] M. Bower, E. Cloud, H. Duvoisin, D. Long, and J. Hackett, Development of Automatic Target Recognition for Infrared Sensor-Based Close Range Land-Mine Detector. Orlando, FL: Martin Marietta, [4] K. Fuelop and J. Hall, Thermal infrared land-mine detection, Defence Science Technology Organization, Salisbury, Australia, Tech. Rep. DSTO-TR-0295 AR , Jan [5] G. Carpenter and S. Grossberg, Pattern Recognition by Self-Organizing Neural Networks, New York: Academic, 1987, pp [6] A. Filippidis, L. C. Jain, and P. Lozo, Degree of familiarity ART2 in fuzzy fusion landmine detection, in Int. Conf. Knowledge-Based Intelligent Electronic Systems (KES 98), Adelaide, Australia, Apr. 1998, pp [7] R. M. Harlick, K. Shunmuhham, and I. Distein, Textural features for image classification, IEEE Trans. Syst., Man, Cybern., vol. SMC-3, Nov [8] P. J. Whitbread, Multispectral texture, Ph.D. dissertation, Univ. of Adelaide, Australia, Oct [9] V. K. Shettigara, S. G. Kempinger, and E. Kruzins, Fusion of texture and intensity characteristics for target detection in SAR images, in Proc. Int. Workshop IAIF 97: Image Analysis and Information Fusion, Adelaide, South Australia, 1997, pp [10] P. P. Ohanian and R. C. Dubes, Performance evaluation for four classes of textural features, Pattern Recognition, vol. 25, pp , [11] Neuro Genetic Optimizer Version Redmond, WA: BioComp. [12] A. Filippidis and L. C. Jain, Identity attribute information in a multi-band aerial photo image using neural networks to automatically detect targets, Int. J. Knowledge-Based Intell. Eng. Syst., vol. 1, pp , Jan Abstract Character recognition systems can contribute tremendously to the advancement of the automation process and can improve the interaction between man and machine in many applications, including office automation, check verification and a large variety of banking, business and data entry applications. The main theme of this paper is the automatic recognition of hand-printed Arabic characters using machine learning. Conventional methods have relied on hand-constructed dictionaries which are tedious to construct and difficult to make tolerant to variation in writing styles. The advantages of machine learning are that it can generalize over the large degree of variation between writing styles and recognition rules can be constructed by example. The system was tested on a sample of handwritten characters from several individuals whose writing ranged from acceptable to poor in quality and the correct average recognition rate obtained using cross-validation was 89.65%. Index Terms Arabic characters, feature extraction, hand-printed characters, inductive logic programming, parallel thinning, pattern recognition, structural classification. I. INTRODUCTION Character recognition is commonly known as optical character recognition (OCR) which deals with the recognition of optical characters. The origin of character recognition can be found as early as 1870 [1] while it became a reality in the 1950 s when the age of computer arrived [2]. Commercial OCR machines and packages have been available since the mid 1950 s. OCR has wide applications in modern society: Document reading and sorting, postal address reading, bank check recognition, form recognition, signature verification, digital bar code reading, map interpretation, engineering drawing recognition, and various other industrial and commercial applications [3] [7]. The products that are currently commercially available for character recognition are limited to the recognition of typed text within a restricted number of fonts, or on-line recognition of hand-written characters. Products to perform off-line hand-printed text recognition are not available, although many approaches have been proposed. In fact there has recently been a high level of interest in applying machine learning to solve this problem [8] [10]. Much more difficult, and hence more interesting to researchers, is the ability to automatically recognize handwritten characters [11]. The complexity of the problem is greatly increased by noise and by the wide variability of handwriting as a result of the mood of the writer and the nature of the writing. Analysis of cursive scripts requires the segmentation of characters within the word and the detection of individual features. This is not a problem unique to computers; even human beings, who possess the most efficient optical reading device (eyes), have difficulty in recognizing some cursive scripts and have an error rate of about 4% on reading tasks in the absence of context [12]. Different approaches covered under the general term character recognition fall into either the on-line or the off-line category, each having its own hardware and recognition algorithms. In on-line character recognition systems, the computer recognizes the symbols as they are drawn [13] [16]. The most common writing surface is the /00$ IEEE Manuscript received November 30, 1998; revised November 12, The author is with the School of Computer Science, University of New South Wales, Sydney 2052, Australia ( amin@cse.unsw.edu.au). Publisher Item Identifier S (00)