Color image recognition by use of a joint transform correlator of three liquid-crystal televisions

Transcription

1 Color image recognition by use of a joint transform correlator of three liqui-crystal televisions Mei-Li Hsieh, Ken Y. Hsu, an Hongchen Zhai We present a joint transform correlator for color image recognition by using three liqui-crystal spatial light moulators. A metho for simultaneously obtaining the correlation peaks of re, green, an blue is propose an experimentally emonstrate Optical Society of America OCIS coes: , , , Introuction Use of the joint transform correlator JTC is one of two common techniques use for optical image recognition. 1 3 In this technique, liqui-crystaltelevision LCTV -base JTC systems have been wiely use for monochromatic signal etection an ientification. 4 In practice, many visual signals are colore. Both the shapes an the colors of the input patterns are essential characteristics for pattern recognition. Thus extening the capability of the optical pattern recognition system for colorimage inputs is a very attractive task. In fact, color pattern recognition systems that use matche filters 5 8 an JTCs 9 11 have been emonstrate. In those systems, either three lasers with ifferent colors are use or an aitional grating evice is use to separate the three original-color re-greenblue RGB components of a composite input image. In both cases, the systems are complicate. In recent years there has been a rapi evelopment in isplay technology. High-resolution LCTV projectors that contain three panels, one panel isplaying one of the three original colors of the input image, are commercially available. In these evices, separation of the RGB components of a composite image is automatically performe by the built-in circuit boar. The color separation an isplaying characteristics of the LCTV projectors M.-L. Hsieh an K. Y. Hsu are with the Institute of Electro- Optical Engineering, National Chiao Tung University, Hsin-Chu, Taiwan, China. H. Zhai is with the Laboratory of Information Science, Institute of Moern Optics, Nankai University, Tianjin 30071, China. Receive 28 April $ Optical Society of America provie a goo application for color image recognition. In this paper a three-channel JTC structure for performing color pattern recognition is presente. In the system, each channel is a JTC correlator responsible for the recognition of a specific color component RGB of the color image. The selection of suitable focal lengths of the lenses of the jointtransforme correlators of RGB color components allows the correlation peaks for RGB components to be isplaye simultaneously at separate locations on a screen, an the color image recognition can be achieve. In the following section, the principle of our system will first be escribe. Then Section 3 presents the system parameters an experimental results. Section 4 gives some conclusions. 2. Principle In our system, three liqui-crystal isplay LCD panels of a LCTV projector are use as the input evices, each panel isplaying a ifferent color component of the input image. The colors of the input pattern are separate into RGB colors by the river circuits of the LCTV projector. If each LCD panel is put at the front focal plane of a Fourier lens, then the Fourier spectrum of each color can be obtaine at the back focal plane. Owing to the pixelate structure of the LCD evice, each Fourier spectrum has multiple iffraction orers. One way to separate the Fourier spectra of color images is to choose ifferent orers of ifferent colors. Then the Fourier spectra of RGB components will be spatially separate. But the correlation peaks for three ifferent colors will appear at the same position on the correlation plane when the focal lengths for ifferent colors are the same. Another way of achieving this is to choose the same iffraction orer but with 1500 APPLIED OPTICS Vol. 41, No March 2002

2 Fig. 1. ifferent focal lengths for ifferent colors. In this paper we use the secon metho. Figure 1 shows our system of the joint transform correlator. The RGB components of input objects an the reference images are jointly isplaye on three LCTV panels. Consier a reference image g x, y an an object image h x, y, both of which are color images. The RGB components of the images are separate by the river circuit of the projector an are isplaye on the three corresponing LCTV panels. The total inputs of the three panels can be written as f in x, y f r x, y f g x, y f b x, y g r x 2, y h r x 2, y t r x, y g g x 2, y h g x 2, y t g x, y g b x 2, y h b x 2, y t b x, y, where f r, f g, an f b represent the patterns that are isplaye on the re, green, an blue LCTV, respectively. t x, y represents the amplitue transmittance of the LCTV panel, an represents the spatial separation between the reference an the object image on the LCD panel. As Fig. 1 shows, each pair of the R, G, B inputs are jointly transforme by lenses L1, L2, an L3, respectively. The summation of the three Fourier spectra can be foun to be F, G r, exp i 2 2 H r, exp i 2 2 T r, Color pattern recognition in a joint transform correlator system. (1) G g, exp i 2 l g 2 H g, exp i 2 T g, G b, exp i 2 2 H b, exp i 2 2 T b,, where R represents convolution operation; an are the coorinates at the Fourier plane; G an H are the Fourier transforms of the patterns g x, y an h x, y ; l r, l g, an l b are the focal lengths of L1, L2, an L3; an T r, T g, an T b are the Fourier transforms of the amplitue transmittance of the re, green, an blue LCTV panels, respectively. Because the LCTV panel is pixelate, its Fourier spectra has multiple iffraction orers an thus T r, T g, an T b can be written as a summation of replica of Fourier spectra, (2) T i, A mn m i, n i, m n i r, g, b, m 0, 1, 2,..., n 0, 1, 2,..., (3) where A mn is the iffracte amplitue of the mn orer an A mn sinc md i sinc nd i an D is the pixel size of the LCTV panels. If we put an aperture at the Fourier plane such that only one iffraction orer passes through, then spatial filtering can be achieve. For example, if we choose the first orer m 1, n 0 of the Fourier spectra, then the 10 March 2002 Vol. 41, No. 8 APPLIED OPTICS 1501

3 joint-transforme spectra in Eq. 2 can be rewritten as F 1, c 1 G r r, exp i 2 2 r H r r, exp i 2 2 r c 2 G g g, exp i 2 g H g g, exp i 2 g c 3 G b b, exp i 2 2 b H b b, exp i 2 2 b, where c 1, c 2, an c 3 are the amplitues of the firstorer iffraction for the re, green, an blue channels, respectively. Note that the three terms in Eq. 4 possess the same spatial orientation on the Fourier planes. But because we choose ifferent focus lengths of the Fourier lenses for the three LCTV, i.e., l r l g l b, then the centers of the three Fourier spectra are locate at three positions, r, g, an b. These Fourier spectra are recombine by beam splitters BS1 an BS2 an are image into camera CCD1 by lenses L5 an L6. CCD1 etects the power spectra of the joint transform, which can be written as F 1, 2 C 1 G r r, 2 H r r, 2 G r r, H* r r, exp i 2 r c.c. C 2 G g g, 2 H g g, 2 G g g, H* g g, exp i 2 c.c. C 3 G b b, 2 l g g H b b, 2 G b b, H* b b, exp i 2 b (4) c.c., (5) where C 1, C 2, an C 3 are the intensities for the re, green, an blue spectra, respectively, an * an c.c. represent the complex conjugate. Note from Eq. 5 that the grating perio of the power spectra for re, green, an blue are, l g, an, respectively. The power spectra etecte by CCD1 is isplaye on LCTV2, an then it is Fourier transforme by lens L4. The output intensity is etecte by CCD2, which can be written as I output 1 F 1, 2 2 C 1 g r g r h r h r C 2 g g g g h g h g C 3 g b g b h b h b 2 C 1 g r h r x l l r C 2 g g h g x l l g C 3 g b h b x l 2 l b C 1 g r h r x l l r C 2 g g h g x l l g C 3 g b h b x l 2 l b, (6) where represents the correlation operation an l is the focal length of lens L4. In Eq. 6, it is seen that the autocorrelation terms of the input an the reference images are overlap an are locate at the center of the output plane an that the cross-correlation peaks for re, green, an blue patterns are spatially separate at six positions, l, l l g, an l, respectively. If the input pattern g x, y is ientical with the reference image h x, y both in shapes an colors, then six correlation peaks will appear on the output plane. If g x, y an h x, y are ientical in shapes an both are re, then there will be only two correlation peaks appearing at the locations l. However, if the two patterns are ientical in shapes but are ifferent colors, e.g., one is re an the other one is green, then there shoul be no correlation peak on the output plane. 3. Experimental Results In our experiment the focal lengths of the lenses for re, green, an blue signals are 60, 84.1, an 40 cm, respectively. The focal length of lens L4 is 60 cm. The focal lengths of lens L5 an L6 are ientical an are equal to 20 cm. Figure 2 shows the input an the reference patterns; both are the white characters A. The separation between the two characters is 0.4 cm on the LCD screen. Therefore the grating perios of the power spectra corresponing to re, green, an blue are 77.1, 108.1, an 51.4 m, respectively. The pixel size of CCD1 is 11 m 13 m. Thus the joint transform power spectra are etectable by CCD1. The first-orer power spectra of the three colors are 1502 APPLIED OPTICS Vol. 41, No March 2002

4 Fig. 2. Original input patterns. selecte by apertures with a iameter of 1 mm an are image onto CCD1 by lenses L5 an L6. The etecte power spectra is shown in Fig. 3. It can be seen that the power spectra of the three colors are separate an locate at three ifferent positions in the Fourier plane. It can also be seen that the top spectra is strongest an the bottom spectra is the weakest because the input image has a large amount of green an a small amount of blue. The etecte power spectra is isplaye on LCTV2 an is Fourier transforme by L4. The correlation output signal is etecte by CCD2 an is shown in Fig. 4. It is seen that there are six cross-correlation peaks prouce by the ientical white patterns. The central bright spot is the autocorrelation peak. The two points nearest to the central peak represent the correlation outputs for the blue color of the input patterns. The next two points represent the correlation outputs for re. An the two points farthest from the central peak represent the correlation outputs for green. With the experimental parameters, the istances between the central peak an the correlation peaks for re, green, an blue are 0.4, 0.6, an 0.29 cm, respectively. Fig. 4. Correlation output for two white patterns of Fig. 2. Fig. 5. Correlation output for white an blue patterns. The left image shows the input pattern of letters A s in ifferent colors. The upper A is blue, an the lower A is white. The right image shows the correlation output. Fig. 6. Correlation output for re an blue input patterns. The upper A is re, an the lower A is blue. The right image shows the correlation output. Fig. 3. Power spectra of two white input patterns. The correlation result of changing the input characters to blue while the reference image is kept white is shown in Fig. 5. It can be seen that only two correlation peaks for blue have been observe. The relative values of the correlation peaks for re, green, an blue are 36, 56, an 255, respectively. Next, the 10 March 2002 Vol. 41, No. 8 APPLIED OPTICS 1503

5 correlation output of changing the color of the input character to re an the reference character to blue is shown in Fig. 6. It can be seen that there is no correlation peak in the correlation plane. 4. Conclusion In this paper we have propose an experimentally emonstrate a joint transform correlation system for color image recognition. The system utilizes three LCD panels to isplay RGB components of a color image. When ifferent focal lengths for Fourier lenses of the RGB images are chosen, the correlation peaks of the three color components are spatially separate an locate at ifferent positions. By etection of the correlation outputs, color image recognition has been achieve. This research is supporte by grants from the National Science Council, Taiwan, uner contract NSC E an from the Ministry of Eucation uner grant 89-E-FA M.-L. Hsieh is a postoctoral research fellow supporte by the Lee & MTI Networking Research Center at National Chiao-Tung University in Taiwan. References 1. A. B. VanerLugt, Signal etection by complex spatial filtering, IEEE Trans. Inf. Theory IT-10, C. S. Weaver an J. W. Gooman, A technique for optically convolving two functions, Appl. Opt. 5, X. J. Lu, F. T. S. Yu, an D. A. Gregory, Comparison of VanerLugt an joint transform correlators, Appl. Phys. B 51, F. T. S. Yu an X. J. Lu, A real-time programmable joint transform correlator, Opt. Commun. 52, N. K. Shi, Color-sensitive spatial filters, Opt. Lett. 3, F. T. S. Yu an T. H. Chao, Color signal correlation etecte by matche spatial filtering, Appl. Phys. B 32, F. T. S. Yu an B. Javii, Experiments on real-time polychromatic signal etection by matche spatial filtering, Opt. Commun. 56, M. S. Millan, J. Campos, C. Ferreira, an M. Yzuel, Matche filter an phase only filter performance in color image recognition, Opt. Commun. 73, F. T. S. Yu, S. Jutamulia, R. V. Yelamarty, an D. Gregory, Aaptive joint transform correlator for real-time color pattern recognition, Opt. Laser Technol. 21, F. T. S. Yu, Z. Yang, an K. Pan, Polychromatic target ientification with a color liqui-crystal-tv-base joint-transform correlator, Appl. Opt. 33, H. Zhai, G. Mu, J. Sun, X. Zhu, F. Liu, H. Kang, an Y. Zhan, Color pattern recognition in white-light joint transform correlation, Appl. Opt. 38, APPLIED OPTICS Vol. 41, No March 2002