Rotation/ scale invariant hybrid digital/optical correlator system for automatic target recognition V. K. Beri, Amit Aran, Shilpi Goyal, and A. K. Gupta * Photonics Division Instruments Research and Development Establishment Dehradun-248 008, INDIA ABSTRACT Optical correlator systems due to their inherent parallel processing capabilities have the potential of processing large data bank to enable rapid identification of targets within an input signal. Purely optical correlation systems are not practical for real time applications. In this paper, we discuss and implement the hybrid correlator system. In this system, the input scene is Fourier transformed digitally and is multiplied with the pre- synthesized digital filter. The product displayed on a spatial light modulator is optically Fourier transformed to obtain correlation peaks. Key words: Target recognition, Hybrid correlator, MACH filter. 1. INTRODUCTION Pattern recognition involves the identification of a reference pattern in a collection of input patterns 1-15. This can be accomplished by performing optical correlation of reference and input patterns and detecting the obtained correlation signal. Most of the 4f optical correlators 1, 2 that are based on frequency domain filtering, display the digitized input target captured using CCD camera on a spatial light modulator (SLM), placed at the first focal plane of a Fourier transforming (FT) lens. A well-collimated coherent beam illuminates the target displayed and its optical Fourier transform falls on the template filters displayed on a second SLM placed at the back focal plane of the FT lens. The optical Fourier transform of this multiplication is captured by a CCD camera to observe correlation peaks 3. There have been constant efforts to develop better filters for identification of objects under various invariances like rotation, scale, variation of illumination, background noise etc. A notable contribution to the field was the simulation of the Synthetic Discriminant Function (SDF) 1,2,4,5 filter. Since the introduction of SDF, a number of significant advances have been documented. These include Minimum Average Correlation Energy (MACE) 6 and Maximum Average Correlation Height (MACH) 7 filters. MACH filter is optimized to detect a target in the presence of rotation/scale distortions. In the hybrid approach 8-11 the input scene is Fourier transformed using fast Fourier transform (FFT) algorithm and multiplied digitally with a pre-synthesized digital filter. This product is displayed over the frequency plane SLM and its optical FT is obtained. This approach has many advantages over the fully optical approach, which has been discussed in the next section. For pattern recognition with real world targets under the effect of scale/rotation, it is necessary to perform several correlations per input signal. This is done by displaying large number of template filters for each input signal which puts constraint of different frame rates requirement on input and frequency plane SLMs as well as on detection plane. To handle large amount of data we have conceptualized hybrid digital/optical correlator architecture 8-11. In this paper, we have discussed digital/optical correlator scheme for real time automatic target recognition. Real world targets have been captured using a CCD camera. The MACH filter was synthesized using a set of training images. The Fourier transform of the input scene is computed digitally using the FFT algorithm. The digital multiplication of target s FT with pre-synthesized rotation/scale invariant MACH 5,7 filter is displayed on Boulder Nonlinear System (BNS) SLM having a frame rate upto 1000 frames per second. This SLM is a ferroelectric liquid crystal based, which works in the reflection mode. The inverse FT of the frequency plane product displayed on the SLM is achieved optically to obtain correlation at fast rate. We have used the lab prototype developed at IRDE Dehradun 12 for implementing this approach. The simulation and experimental results for a tank type of target have been presented. * Corresponding author. E-mail: akgupta@irde.res.in, Tel. + 91-135-2787089, Fax: + 91-135-2878161
1.1 MAIN ADVANTAGES OF THE HYBRID APPROACH The followings are the advantages of the hybrid digital/optical correlator proposed by Young and his co-workers 8-11 over the fully optical correlator based on VanderLugt geometry. 1. The laser source requirements are much reduced from those necessary in an all optical correlator, particularly the power level since the only major losses are those associated with the frequency plane phase SLM. There is no particular need for high frequency stability but it is important that the laser should emit in TEM 00 so that a Gaussian profile is maintained. 2. The input SLM is avoided and a precise spectrum is available for subsequent digital or optical mixing with the stored templates. 3. There is no requirement to match, and maintain a match, of an optically formed spectrum to the frequency plane SLM as in an all-optical correlator, instead the SLM being illuminated with an apodized plane wave directly from the laser source. 4. It is now possible to compute a two-dimensional Fourier transform at video-rates (25 Hz) and video resolution (512 512 pixels) with various modern digital signal processing chip sets. 5. The optical design and alignment is much simplified from that of an all optical correlator since there is requirement to perform only a single optical Fourier transform between the single SLM and a high pixel density CCD array. 6. Lens design requirements are complex. It is possible to design a relatively simple FT lens system, which has the correct output scaling, without an excessive focal length being necessary. 7. Reduced mechanical alignment difficulties. 8. Improved robustness of the device to external mechanical disturbance. 2. SIMULATION STUDIES AND RESULTS In order to perform simulation study images of two different tanks were used. For rotation, a set of training images were formed by rotating tank1 (true class) in plane at an interval of 3 0, whereas for scaling a set of true class images with a 5% scale variation was formed. Tank2 was used as false class image. Targets used for filter synthesis and the synthesized filter are shown in Fig.1 Tank 1 (a) Tank 2 (c) Synthesized MACH filter (d) Set of training images (b) Fig. 1 Simulation studies were performed on Matlab platform. From the results it can be observed that the peak height in cross correlation is less than the peak height for same class of images. The MACH filter synthesized was multiplied with the digital FFTs of various targets and the product was inverse Fourier transformed to observe the correlation peaks. Simulation study results are shown in Figs. 2-3 for trained, untrained, and cross targets at different scale and rotation orientations.
(a) (b) (c) Fig. 2 Simulation results for scale invariant hybrid correlator (a) original target (b) trained target with 85%scale (c) trained target with 124% scale (d) cross target (e) untrained target with 83% scale and (f) untrained target with 126% scale (a) (b) (c) Fig. 3 Simulation results for rotation invariant hybrid correlator (a) original target (b) trained target with 27 o clockwise rotation (c) trained target with 27 0 anticlockwise rotation (d) cross target (e) untrained target with 22 o clockwise rotation and (f) untrained target with 22 0 anti clockwise rotation. 3. EXPERIMENTAL STUDIES AND RESULTS The schematic of experimental set-up used is shown in Fig. 4. The Photonics correlator developed at IRDE is of size 33 17.5 90 cm 3 and makes use of a single 512 512 pixels (with pixel size 7 7 µm) BNS SLM for displaying the product. A narrow beam of light from a 5 mw Laser diode operating at wavelength 670 nm is used as a coherent light
source. The collimated beam after passing through a polarizing beam splitter illuminates the multiplication product displayed on the SLM. The optical FT is obtained with an FT lens of focal length 135 mm. A Sony XC-ST-30 CCD camera placed at the back focal plane of the FT lens captures the correlation signal at video rates. Experimental results are shown in Fig.5 for trained, untrained and cross targets at different scale and rotation orientations. PBS SLM SLM DRIVER LASER SF CL FT LENS PC CCD Set of trained images MACH Filter Multiplication Digital Domain Real time target Digital FFT Fig.4. Schematic of hybrid correlator. SF: spatial filter, CL: collimating lens, PBS: polarizing beam splitter, SLM: spatial light modulator, PC: personal computer, CCD: charge coupled device. (a) (b) (c) Fig. 5 Experimental results for scale/rotation invariant hybrid correlator (a) original target (b) trained target with 85% scale (c) untrained target with 83% scale (d) cross target (e) trained target with 27 o clockwise rotation (f) untrained target with 22 0 anti clockwise rotation
4. CONCLUSION We have studied the hybrid digital/optical correlator scheme using MACH filter for scale/ rotation invariance. Our study verifies that the synthesized MACH filter could achieve rotation invariance from 30 to 30 and scale invariance from 75% to 140%. We plan to undertake study of the combined effect of scale and rotation invariance in future at a much faster rate. ACKNOWLEDGEMENT The authors are grateful to Shri J. A. R. Krishnamoorty, Director, IRDE Dehradun for his encouragement and giving permission to present and publish this paper in the 7 th International Conference on Photonics 2004. REFERENCES 1. J. W. Goodman, Introduction to Fourier Optics, 2 nd Ed., Mc-Graw Hill, NewYork, 1996. 2. F. T. S. Yu and S. Jutamulia, Optical Pattern Recognition, Cambridge Univ. Press, NewYork, 1998. 3. B. Javidi, Ed., Image Reconstruction and Classification; Algorithms, Systems, and Applications, Marcel Dekker, Inc., New York, 2002. 4. G. S. Pati and K. Singh, Discrimination-sensitive rotation-invariant joint transform correlator using gradient preprocessing and circular harmonic components, Opt. Eng. 37, 43-48, 1998. 5. B. V. K. Vijaya Kumar, Tutorial survey of composite filter designs for optical correlators, Appl. Opt. 31, 4773-4801, 1992. 6. A. Mahalanobis, B. V. K. Vijaya Kumar, and D. Casasent., Minimum average correlation energy filters, Appl. Opt. 26, 3633-3640, 1987. 7. A. Mahalanobis, B. V. K. Vijaya Kumar, S. Song, S. R. F. Sims, and J. F. Epperson, Unconstrained correlation filters, Appl. Opt. 33, 3751-3758, 1994. 8. R. Young, C. R. Chatwin, and B. R. Scott, High speed hybrid optical/digital correlator system, Opt. Eng. 32, 2608-2614, 1993. 9. R. Young, S.-P. Huang, G. Li, T. Koukoulas, D. Budcett, and C. Chatwin, Implementation and performance considerations of hybrid digital/optical correlator configuration, Proc. SPIE 3715, 16-31, 1999. 10. P. M. Birch, F. C. Tournier, D. Budgett, R. Young, and C. Chatwin, Optical and electronic design of a hybrid digital-optical correlator system, Opt. Eng. 41, 32-40, 2002. 11. P. M. Birch, G. Li, F. C. Tournier, R. Young, D. Budgett, and C. Chatwin, Optical design of a miniature Fourier transform lens system for a hybrid digital-optical correlator, Opt. Eng. 41, 1650-1654, 2002. 12. V. K. Beri, A. K. Musla, D. P. S. Negi, and A. K. Gupta, Design and development of a compact portable joint transform correlator, Proc. Int l Confer. on Fiber Optics & Photonics (PHOTONICS 2000), Calcutta, India, Dec. 18-20, 2000, Eds. S. K. Lahiri, R. Gangopadhyay, B. K. Mathur, A. K. Dutta, S. K. Ray, and S. Das, pp. 630-632, 2000. 13. B. V. K. Vijaya Kumar, A. Mahalanobis, and A. Takessian, Optimal tradeoff circular harmonic function correlation filter methods providing controlled in-plane rotation, IEEE Trans. Image Process. 9, 1025-1034, 2000. 14. A. Mahalanobis, R. R. Muise, and S. R. Stanfill, Quadratic correlation filter design methodology for target detection and surveillance applications, Appl. Opt. 43, 5198-5205, 2004. 15. M. S. Alam and M. A. Karim, Advances in pattern recognition algorithms, architectures, and devices, Special Section Guest Editorial, Opt. Eng. 43, 1702-1704, 2004.