Image stitching and image reconstruction
of intestines captured using radial
imaging capsule endoscope
Mang Ou-Yang
Wei-De Jeng
Yin-Yi Wu
Lan-Rong Dung
Hsien-Ming Wu
Ping-Kuo Weng
Ker-Jer Huang
Luan-Jiau Chiu
Wei-De Jeng Yin-Yi Wu
National Chiao-Tung University
Department of Electrical and Control Engineering Hsinchu City 30010, Taiwan
E-mail:[email protected] Lan-Rong Dung
National Chiao-Tung University Department of Electrical Engineering Hsinchu City 30010, Taiwan
Hsien-Ming Wu Ping-Kuo Weng Ker-Jer Huang Luan-Jiau Chiu
Chung-Shan Institute of Science and Technology Lung-Tan 90008-8-8, Taiwan
(mean absolute error, mean square error, Pearson correlation coefficient, and deformation processing) are used to stitch the images together. The Pearson correlation coefficient method is the most effective algorithm because it yields the highest peak signal-to-noise ratio, higher than 80.69 compared to the original image. Furthermore, a living animal experiment is carried out. Finally, the Pearson correlation coefficient method and vector deformation processing are used to stitch the images that were captured in the living animal experiment. This method is very attractive because unlike the other methods, in which two lenses are required to reconstruct the geometrical image, RICE uses only one lens and one mirror.© 2012 Society of Photo-Optical Instrumentation Engineers (SPIE). [DOI:10.1117/1.OE.51.5.057004]
Subject terms: radial imaging capsule endoscope; mean absolute error; mean square error; pearson correlation coefficient; peak signal-to-noise ratio; stitch images; correlation.
Paper 111503 received Dec. 1, 2011; revised manuscript received Feb. 29, 2012; accepted for publication Mar. 1, 2012; published online May 4, 2012.
1 Introduction
Diseases of the intestine are very difficult to treat because the intestine is very long and winding. Although the colono-scopy has been available for many years, it is still very difficult to use to obtain internal images because the length of intestine is about 7 m. Therefore, the Given Image Co. (Yoqneam, Israel) proposed a first-generation capsule endoscope called the M2A capsule endoscope.1,2 After
this new endoscope entered the market, several companies and laboratories began conducting relevant research. They included a radio frequency (RF) system laboratory, Olympus in Japan,3 and Intelligent Microsystems in Korea, China, and Taiwan.4,5 The capsule endoscope, developed several
years ago, solved the main problems of capturing images inside the intestine.6 Unfortunately, this kind of capsule endoscope can take only the front images; therefore, it also is known as front imaging capsule endoscope (FICE).7 However, about 50,000 images are captured
during a typical period of treatment, and having a doctor examine these images one at a time takes a long time. There-fore, this work develops a new generation of radial imaging capsule endoscopes (RICE) with a different imaging method, in which a cone mirror is used to obtain radial images.7 In
the RICE system, the regions in the images that are captured at different times have overlap, as shown in Fig. 1. The image from the interior toward the exterior, black, green, white, and blue in concentric rings when t is zero. Then, RICE moves forward to t¼ 1; the image then becomes
green, white, and blue in concentric rings from the interior toward the exterior; similarly at t is 2 and 3. From this exam-ple, RICE provides images that can be related to each other, enabling the complete image to be reconstructed.7 The images can be combined using a mathematical algorithm, enabling doctors to diagnose a disease more effectively because they do not need to examine these images one at a time.
However, some problems are encountered with the RICE system. They fall into two categories: imaging and lighting. With respect to the former, the cone mirror increases aberra-tion and compresses the image, reducing the image resolu-tion.7The structure of the RICE system differs from that of the FICE system. In the FICE system, the problem of stray light is solved by putting the LEDs on the elliptical focal plane of the dome, because of the stationarity optical path length.8 However, the RICE system is composed of an extra cone mirror, so the problems of stray light and over-blooming are very serious. Accordingly, this work mainly focuses on stitching the sequential images using different algorithms and solving the lighting and imaging problems by image processing. Finally, the image is reconstructed after image stitching.
2 Radial Imaging 2.1 RICE Component
The greatest difference between RICE and FICE is in the imaging: the RICE system uses a cone mirror, but the other components, such as the imager, illuminator, viewing window, and optical module, do not differ much from those
in the FICE system.7 Table 1 compares RICE with FICE; RICE uses a cone mirror and lenses to capture radial images, and the CCD resolution and pixel size are the same compared to FICE.
2.2 Radial Imaging Processing
This section introduces image processing in the RICE sys-tem. Figure2shows the steps. First, the RICE passes through a cylindrical object and takes photographs at different times. Second, the images are“unwrapped” by performing a coor-dinate transformation. Finally, an algorithm is utilized to stitch the unwrapped images and construct geometrical image of the object.
3 Analysis of RICE Imaging Properties 3.1 Compression of Images
The geometry of the cone mirror causes the problem of image compression in the RICE system. Each image is severely compressed when close to its center. This effect is also called warping, and it is shown in Fig.3. Clearly, images obtained using the RICE system can be treated as having polar coordinates. They are unwrapped by a coordinate transformation, using equations such as Eqs. (1) and (2).9 In Fig. 3, fðx; yÞ represents a warped image; Oðx0; y0Þ is
Fig. 1 The images have overlapping regions in the RICE system.
Table 1 The configuration comparison between FICE and RICE.
Item
Capsule
FICE RICE CCD resolution 512 × 512 512 × 512 Pixel size 5.6μm 5.6μm LED White× 6 White× 6 Imaging region Front region Side region
(radial imaging) Window type Semispherical
type
Annular type Imaging
component
Doublet lens Doublet lens and cone mirror Cone mirror size None 3.35 mm
Fig. 2 The image-processing steps of the RICE system.
Fig. 3 The top part reveals that the images are compressed by cone mirror; the bottom part shows the image is unwrapped by coordinate transform.
Fig. 4 A lighting problem in RICE causes over-blooming and low lighting uniformity.
the center of the warped image, and gðθ; zÞ is the unwrapped image obtained by applying Eqs. (1) and (2) to the warped image. z¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðx − x0Þ2− ðy − y0Þ2 q (1) θ ¼ tan−1 y− y0 x− x0 . (2) it is unwrapped.
3.3 Image Stitching Algorithms
Image stitching has been developed over a long period of time. It is always used in panoramic systems, which take a sequence of mutually related photographs at different times. RICE can also be applied as a panoramic system, which requires image stitching. The correlation between two sets of data can be computed mathematically using several methods. These include feature-based robust estimation,11the multi-resolution method,12scale-invariant feature transforma-tion (SIFT),13statistical methods, and energy maps.14In this study, three statistical algorithms are used to stitch images in the RICE system. These are the mean absolute error (MAE), mean square error (MSE) and Pearson correlation coefficient-based methods. The corresponding formulas are Eqs. (3) through (5), respectively. The MAE uses an error to describe the difference between two sets of data. If the value is low, then the sets of data are highly correlated. The MSE method is based on the same concept, but the value depends on the square of the difference between data.15–18The Pearson cor-relation coefficient Cðx; yÞ is a statistic that is obtained by computing the covariance and variance of data sets, a value of close to unity means a strong correlation.19,20
object sample.
Fig. 6 (a) The clear imaging region of warped and (b) unwrapped images.
MAE ¼ 1 m× n X m×n−1 i¼0 jSðiÞ − RðiÞj (3) MSE ¼ 1 m× n X m×n−1 i¼0 jSðiÞ − RðiÞj2 (4) Cðx; yÞ ¼ P m P n½Sðm; nÞ − us½Rðm − x; n − yÞ − uR ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P m P n ½Sðm; nÞ − us 2 r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP m P n ½Rðm − x; n − yÞ − uR 2 r . (5) S is a sample image; R is the reference image; m× n is image size; x and y are the points of the highest correlation between S and R computed from the Pearson correlation coefficient, and μSandμRare the average of the sample and reference images.
Hence, when the highest position of the highest correlation between the sample and reference images is determined, the images can be stitched by shifting the sample image to this position and the reference and sample images then combined.
4 Experiments and Image Processing 4.1 Experimental Setup and Image Capture
An experiment is carried out using RICE to capture radial images. The object is a tube, to mimic an intestine. The left side of Fig.5shows the experimental devices, including the CCD, lenses, the cone mirror, and a simulated object. The simulated object is a paramecium figure, which shown on the right side of Fig.5.7A paramecium figure is chosen because
it has a complex shape with many edges, and so can be used to test whether the image-stitching algorithm is strong. A triaxial platform is utilized to move the RICE, and the shot rate is one image per millimeter. The RICE scanning of the object yields 70 images, which can be processed using the above algorithm.
Before the images can be stitched, they must be unwrapped using Eqs. (1) and (2). The Oðx0; y0Þ of the warped image is (241, 241) because this point corresponds to the optical axis in the image plane. Figure6shows the warped and unwrapped images. Clearly, over-blooming and low resolution are evident at the center of the warped image. After the image is unwrapped, the over-blooming and low resolution are con-centrated in the top region. The red dashed line indicates a clearly imaged region. According to the optical simulation, this region corresponds to the fifth and sixth fields, which had smaller spot sizes than the other fields. Therefore, the RICE imaging system can improve image quality by proces-sing because it can use the high-image-quality region for stitching.7 The arrowhead points from the fifth field to the sixth field. Figure 7 shows some sequential images, both warped and unwrapped.
4.2 Image Stitching
After the images have been unwrapped, the images are stitched using MAE, MSE, and Pearson correlation coefficient–based methods. Figure8shows the image-stitch-ing process. The reference and sample images are two serial images, which have overlapping regions and are shifted relative to each other by an angle that is determined by the rotation of the RICE. Therefore, the three algorithms can be used to compute the position of highest correlation; the sample image is shifted to this position, and then the clearly imaged region is reversed and stitched to the bottom of the reference image. This action yields a new combined image, which includes parts of the reference and sample images.
Fig. 8 The steps of image stitching.
The computing resources used to process the images are an Intel Core 2 Quad CPU, a ASUS P5Q Q9500 mother-board, and 4 GB RAM. Figure9 shows the object sample and the images that are reconstructed from the unwrapped images using the three algorithms. The three reconstructed images are very similar—they cannot be distinguished by the human eye. Their difference is determined using the peak signal-to-noise ratio (PSNR) formula, Eq. (6), in which the denominator MSE is as described above. The object sample and reconstructed image are used here, and the computation of PSNR transforms the colored image to the gray level to eliminate color distortion, which otherwise may affect the PSNR. Table 2 shows the results obtained using the three algorithms. The scan range is the block size of the sample image, which is used to scan the reference image so the position of highest correlation can be deter-mined. The first four data in Table 2 were determined between two serial images. The results indicate that only the Pearson correlation coefficient algorithm is strong
Pearson correlation coefficient algorithm to reconstruct the images, which were captured in an experiment using a living animal. PSNR ¼ 10 × log10 2552 MSE . (6)
4.3 Experiment Using a Living Animal
In the preceding section, an experimental environment that was ideal for image processing was established, enabling some otherwise problematic issues to be neglected. To model the practical use of RICE in the intestine, the ideal experiment should be replaced by an experiment using a living animal. The animal experiment involved a pig that had been raised in the white room of the Animal Technology Institute Taiwan. The pig was about two months old. Its belly was cut open by a professional surgeon to obtain the duode-num. Figure10shows the setup for the animal experiment. The advantage of this animal experiment is that the RICE can move through the duodenum naturally because the living duodenum moves in a manner that pushes the RICE forward. Figure11shows the image reconstruction result of the duo-denum by the Pearson correlation coefficient algorithm. It can be found that misalignment occurred because the best scan range is difficult to find, as the results of Table2 indi-cate. The size of the scan range affects the image stitching results very seriously in the non-rigid world. Besides, the real intestine is not an ideal cylindrical shape, and the surface
Table 2 The image processing results of three algorithms. Data
Algorithm Scan range Time (s) Position Value PSNR Pearson correlation cofficent 13 × 757 6.9 (217,1) 0.957 80.75
12 × 757 6.5 (217,1) 0.957 80.74 11 × 757 5.6 (217,1) 0.957 80.69 MAE 13 × 757 1.9 (217,1) 4.65 80.54 12 × 757 1.8 (217,1) 4.65 80.52 11 × 757 1.5 (218,1) 4.61 79.82 MSE 13 × 757 11.2 (217,1) 37.64 80.63 12 × 757 10.8 (217,1) 37.37 80.50 11 × 757 9.5 (218,1) 37.11 80.21 Fig. 9 (a) The simulated object sample, and image stitching after
image processing by the (b) Pearson correlation coefficient, (c) MSE, and (d) MAE algorithms.
is very rough. Hence, the optical transverse magnification is different in each field when RICE passes through the real intestine. Therefore, as mentioned previously, the image-stitching method should be improved in the non-rigid case. Figure 12 shows the deformation processing by using the vector comparison method. First, we used the Harris corner detector21to find the feature points of the two images S and T so that the feature points can form two vectors, each with a different intensity and direction. Therefore, the vectors should be tuned to the same intensity and direction. By computing the difference between the vectors, the gain of the x and y com-ponent between two images can be determined by Eq. (7),
V1x
V2x¼ Gx
V1x
V2x¼ Gy; (7)
where Vxand Vyare the lengths of the vectors in the x and y
directions, respectively. Once the gain of the deformation points between two images is computed, it can be calibrated, which improves the misalignment problem. Figure13shows
the stitching results between the Pearson correlation coeffi-cient and vector comparison methods, and Fig. 14 shows the image stitching results (each result is stitched with 500 images). From the results, the misalignment problem can be improved by the vector comparison method and is better aligned than Pearson’s method. However, Fig. 14(c) shows that some misalignment still occurred by using the vector com-parison method. This is due to the RICE moving too fast in the intestine, which resulted in losing some image information and caused an unmatchable problem between two sequential images. This problem may be solved by increasing the camera frame rate. Another issue that should be dealt with is intensity inconsistency, which is due to light scattering by the liquid in the intestines, and produces the seam in the stitching result. This issue should be researched in more detail in the future.
4.4 Image Reconstruction in Cartesian Coordinates This section discusses image reconstruction in the RICE system. Some fundamental concepts in geometrical optics are used to determine the geometrical positions. Figure12
shows the positions of the object and the image, where point P1is the object point. To describe geometrical relations
in RICE, Cartesian coordinates (r,θ, z) are used. Points v1
and v2are imaged by the cone mirror at different times. Point
I1and I2are the final images obtained by the lenses; SOand
SIare the object and image distance; d is the air gap between
the tip of the cone mirror and the object principal plane of the lenses; and ROand RIare the cone mirror radius of the object
and image, where RI also equals the image circle in the
image plane. The terms H1 and H2 are the heights of I1
and I2, zðt1Þ and zðt2Þ are the heights of v1 and v1, M is
the transverse magnification, and L is the distance between t1and t2. The image plane in Fig.15shows the image point
and RI. The height and radius of the image plane can be
determined from the given pixel size. Therefore, the transverse magnification M equals RI∕RO, and SI is about 2.86 mm.7
Furthermore, the two similar triangles in Fig. 15 are used to get Eq. (8):
Fig. 12 Vector comparison method steps. (a) Using the Harris corner detector to find the matching points and connect the match points to form the vectors, (b) the corner is determined by a Harris corner detec-tor block and (c) compared with the length and direction with the two vectors.
Fig. 13 The alternative methods of stitching images. (a) Pearson correlation coefficient; (b) vectors comparison method.
Fig. 10 The environment of the animal experiment.
Fig. 11 The image stitching of a young swine duodenum by the Pearson correlation coefficient.
SI r1þ d¼ H1 zðt1Þ¼ H2 zðt2Þ (8)
where zðt1Þ equals z1because the angle between the cone
mir-ror and the z axis is 45 deg, and zðt2Þ equals z1-L. Rewriting
Eq. (8) yields Eq. (9), which can be solved for the depth r1and
position z1 of the object.
H1r1− SIz1 ¼ −H1d H2r1− SIz1 ¼ SIL− H2d ⇒ r1¼ SIL H2− H1− d; z1¼ H1L H2− H1: (9)
That’s the reason why RICE can use only one cone mirror and lenses to reconstruct the object. Figure 16 shows the image reconstruction of duodenum by Eqs. (8) and (9). The left side of this figure displays the full view of the stitch-ing image with young swine duodenum, and the right side zooms in the rectangular range to show the depth information clearly.
Finally, the images captured by FICE and RICE should be compared. Figure17shows the intestine images captured by
FICE and RICE. Figure17(a) shows four images captured by FICE. Due to the depth of focus, the FICE can create clear images in a small area. On the other hand, the RICE takes the radial direction images, and the object distance is short. Hence, it doesn’t need a long depth of focus, and after the clear parts of the images are selected to stitch, the image quality is better, as shown in Fig.17.7,22
5 Discussion and Conclusions
In this work, images captured by the RICE system were processed and reconstructed using MATLAB software (MathWorks, Natick, MA). First, the image compression problem that is caused by the cone mirror is considered. It is solved by transforming the polar coordinates to rectan-gular coordinates. The second step is to identify a clear region in images to find the position of highest correlation between serial images, and then to stitch the images together. This work presented three algorithms for finding the position of highest correlation, involving the MAE, MSE, and Pearson correlation coefficient. An experiment under ideal conditions was carried out to compare these three methods. The method based on the Pearson correlation coefficient was the best: the position of highest correlation was independent of the scan range. The third step was to test the Pearson cor-relation coefficient-based method in a more complex case (in real intestines); thus, an experiment was carried out on a liv-ing animal. It was found that some misalignment occurred without any deformation processing. Hence, the vector com-parison method was used to improve it. Finally, the image
Fig. 17 The intestine images captured by (a) FICE and (b) RICE (after image stitching).
Fig. 15 The imaging relationship between two sequential images.
Fig. 16 Image reconstruction of duodenum (a) full view and (b) the zoom-in view of the block.
Fig. 14 The image stitch results with different patterns. Each pattern is stitched with 500 images. The images (a), (b), and (c) are stitched by using the vector comparison method. (d) The zoom-in view of the blue block in part (a); due to image information loss and intensity inconsistency, it does not look natural.
was reconstructed using Eq. (9). All the object points were obtained using this equation, even when the RICE system had only one lens.
This work presents a preliminary image reconstruction algorithm. However, the reconstructed image that resulted includes some color distortion that is caused by color aberra-tion, so future work must try to figure out how to compensate for this distortion by calibration. This calibration involves estimating the spectral sensitivity of the CCD sensor using eigen-spectrum methods. The estimated sensor spectral sen-sitivity can be mapped to real colors using the standard color mapping function of CIE 1931.
Acknowledgments
This work is supported by the National Science Council of Taiwan under contracts NSC 2220-E-009-033, 100-2218-E-039-001, and NSC 100-2623-E-009-006-D, and the “Aim for the Top University Plan” of National Chiao Tung University and the Ministry of Education, Taiwan. The authors would also like to thank the Chung-Shan Insti-tute of Science and Technology, Delta Electronics Incorpora-tion, and Dr. Jyh-Hung Lin of the Animal Technology Institute Taiwan, who provided experimental assistance and useful information. Finally, the authors are grateful to the instructor of technical writing, Ted Knoy, who helped them edit the manuscript for grammatical and stylistic writ-ing errors.
References
1. Given imaging, “M2A given diagnostic system,” http://www .givenimaging.com/.
2. G. D. Meron, “The development of the swallowable video capsule (M2A),”J. Gastrointest. Endosc.52(6), 817–819 (2000).
3. C. Gheorghe, R. Iacob, and I. Bancila,“Olympus capsule endoscopy for small bowel examination,” J. Gastrointest. Liver Dis. 16(3), 309–313 (2007).
4. A. Uehara and K. Hoshina, “Capsule endoscope NORIKA system,”
Minim. Invasive Ther. Allied Tech.12(5), 227–234 (2003).
5. T. Nakamura and A. Teran,“Capsule endoscopy: past, present, and future,”J. Gastroenterol.43(2), 93–99 (2008).
6. T. Kav and Y. Bayraktar, “Five years’ experience with capsule endoscopy in a single center,” World J. Gastroenterol. 15(16), 1934–1942 (2009).
7. M. Ou-Yang and W. D. Jeng,“Design and analysis of radial imaging capsule endoscope (RICE) system,” Opt. Express19(5), 4369–4383 (2011).
8. V. N. Mahajan,“Optical imaging and aberrations, part i: ray geometrical optics,” Chapter 1, in Gaussian Optics, SPIE Press, Bellingham, WA p. 7 (1998).
9. L. Spacek,“A catadioptric sensor with multiple viewpoints,” J. Robot. Autom. Syst. 51(1), 3–15 (2005).
10. W. D. Jeng et al.,“Design of illumination system in ring field capsule endoscope,”Proc. SPIE7893, 78930E (2011).
11. L. Yao and M. Lizhuang,“A fast and robust image stitching algorithm,” in Proc. of 6th World Congress on Intelligent Control and Automation, Dalian, China, pp. 9604–9608 (2006).
12. X. Fang et al.,“New multi-resolution image stitching with local and global alignment,” J. IET Comp. Vis. 4(4), 231–246 (2010). 13. Q. Zhu and K. Li,“Image stitching using simplified SIFT,” in Proc. of
Information and Automation, IEEE International Conference, Harbin, pp. 1134–1137 (2010).
14. T. Yu and J. Huiyan,“Highly efficient image stitching based on energy map,” Proc. of the 2009 2nd International Congress on Image and Sig-nal Processing, Tianjin, China, pp. 1–5 (2009).
15. S. K. Bar-Lev, B. Boukai, and P. Enis,“On the mean squared error, the mean absolute error and the like,”Comm. Stat.—Theory Methods28(8), 1813–1822 (1999).
16. D. M. Allen,“Mean square error of prediction as a criterion for selecting variables,” Technometrics 13(3), 469–475 (1971).
17. C. J. Willmott and K. Matsuura,“Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance,”Clim. Res.30(1), 79–82 (2005).
18. H. C. Huang and Y. P. Hung,“Adaptive early jump-out technique for fast motion estimation in video coding,” J. Graph. Mod. Image Proc. 59(6), 388–394 (1997).
19. L. Egghe and L. Leydesdorff,“The relation between Pearson's correla-tion coefficient r and Salton's cosine measure,”J. Am. Soc. Inform. Sci. Tech.60(5), 1027–1036 (2009).
20. Y. Wang and L. P. Xu,“A global optimized registration algorithm for image stitching,” in Proc. of 1st International Congress on Image and Signal Processing, Sanya, China, pp. 525–529 (2008).
21. C. Harris and M. J. Stephens,“A combined corner and edge detector,” Proc. Alvey Vision Conf. 147–152 (1988).
22. M. Ou-Yang et al.,“Optimizing the depth of field for short object dis-tance of capsule endoscope,”Proc. SPIE6859, 68591Q (2008).
Mang Ou-Yang received a BS degree in con-trol engineering in 1991, and MS and PhD degrees in electro-optical engineering, in 1993 and 1998, from National Chiao Tung University, Hsinchu, Taiwan. He worked for the Precision Instrument Development Cen-ter, National Science Council of Taiwan, and was in charge of optical metrology for space application from 1994 to 2000. There-after, he worked for Klaser Technology Co. as the leader of the R&D group for develop-ing projection display technology. Since 2004, he has been at the Institute of Optical Sciences, National Central University, Jhongli, Taiwan, as an assistant professor. His research interests are related to optoelectronics industrial instrumentation development, including biomedical optics, microthermal sensors, readout electronics, and projection display technology.
Wei-De Jeng was born in Taiwan in 1986. He received a BS degree from the Department of Electrical Engineering from National Taiwan Ocean University, Taiwan, in 2008 and an MS degree in the Department of Optics and Photonics Engineering from National Central University, Taiwan, in 2010. He is currently studying toward a PhD in the Department of Electrical and Control Engineering at the National Chiao-Tung University. His research interest is the biomedical field.
Yin-Yi Wu received an MS degree in electro-nic and information from the National Yunlin University of Science and Technology, Taiwan, in 1999, and is currently pursuing a PhD degree in electrical engineering at National Chiao Tung University, Hsinchu, Taiwan. He joined the Materials and Elec-tro-Optics Division at the Chung-Shan Insti-tute of Science and Technology. His current interests include integrated-circuit design and image processing.
Lan-Rong Dung (SM’93–M’97) received a BSEE degree from Feng Chia University, Taiwan, in 1988; an MS degree in electronics engineering from National Chiao Tung Uni-versity, Hsinchu, Taiwan, in 1990; and a PhD in electrical and computer engineering from the Georgia Institute of Technology, Atlanta, in 1997. From 1997 to 1999, he was with Rockwell Science Center, Thou-sand Oaks, CA, as a member of the techni-cal staff. He joined the faculty of National Chiao Tung University, Taiwan, in 1999, where he is currently an associate professor in the Department of Electrical and Control Engineering. His current research interests include very large scale integrated design, digital signal processing, hardware-software codesign, and system-on-chip architecture. Professor Dung received the Best Student Award from Feng Chia University, Taiwan, in 1988. He received the VHDL International Outstanding Dissertation Award Celebration in Washington, D.C., in 1997. He is a member of the IEEE Circuits and Systems and Signal Processing societies.
Ping-Kuo Weng earned his BS degree in nuclear engineering from the National Tsing Hua University of Taiwan in 1985. In 1992, he received his PhD degree at the Institute of Electro-Optical Engineering, National Chiao Tung University. Since then, he has been working at the Chung-Shan Institute of Science and Technology in the Solid-State Devices Section. His current research interests include CMOS image sensor design, cameras for medical imaging systems, and system-on-chip architecture.
She received a BS degree at the Department of Computer Science from Feng-Chia Univer-sity, Taiwan, in 1982, and an MS degree in the Department of Electrical Engineering and Computer Science from Yuan-Ze Univer-sity, Taiwan, in 1998. She worked as a research assistant at the Chung-Shan Insti-tute of Science and Technology beginning in 1984. In 1998, she became an assistant research fellow in medical region. Her research interest focuses on image processing, especially in the bio-medical field.
