An Improved Ultra-Low-Power Subsample-based Image Compressor for Capsule Endoscope

An Improved Ultra-Low-Power Subsample-based

Image Compressor for Capsule Endoscope

Meng-Chun Lin

Hsinchu, Taiwan

Email: asurada.ece90g@nctu.edu.tw

Lan-Rong Dung

Hsinchu, Taiwan

Email: lennon@faculty.nctu.edu.tw

Ping-Kuo Weng

Chung-Shan Institute of Science and Technology Solid-State Devices Section

Lung-Tan, Tao-Yuan, Taiwan Email: lennon@nctu.edu.tw

Abstract— The objective of this paper is to further reduce the

power dissipation of the GICam image compressor for capsule endoscope or swallowable imaging capsules. In order to extend the battery life of capsule endoscope, we firstly attempt to analyze the energy distribution and variation of DC/AC coefficients in 2D-DCT domain for twelve testing GI images. According to the analysis result, we can efficiently make use of the subsample technique to reduce the memory requirements of G1, G2 and B components and further propose an improved ultra-low-power subsample-based GICam image compressor, called SGICam, to reduce the power dissipation of compression process. Simulation results have been shown that the SGICam image compressor can significantly save 38.5% power dissipation than GICam image one and the average PSNR is 32.18 dB , while the compression ratio can be as low as 4:1.

I. INTRODUCTION

Gastrointestinal (GI) endoscopy has been popularly applied for the diagnosis of diseases of the alimentary canal includ-ing Crohn’s Disease, Celiac disease and other malabsorption disorders, benign and malignant tumors of the small intestine, vascular disorders and medication related small bowel injury. There exist two classes of GI endoscopy; wired active en-doscopy and wireless passive capsule enen-doscopy. The wired active endoscopy can enable efficient diagnosis based on real images and biopsy samples; however, it causes patients discomfort and pain to push flexible, relatively bulky cables into the digestive tube. To relief the suffering of patients, wireless passive capsule endoscopes are being developed worldwide [1], [2], [3], [4]. The capsule moves passively through the internal GI tract with the aid of peristalsis and transmits images of the intestine wirelessly. The state-of-the-art is the commercial wireless capsule endoscope product, the PillCam capsule, developed by Given Imaging Ltd. The PillCam capsule transmits the GI images at the resolution of 256-by-256 8-bit pixels and the frame rate of 2 frames/sec (or fps). The PillCam has been successfully utilized to diagnose diseases of the small intestine and alleviate the discomfort and pain of patients.

However, based on clinical experience; the PillCam still has some drawbacks. First, the PillCam cannot control its heading and moving direction itself. This drawback may cause image oversights and miss a disease. Second, the resolution of demosaicked image is still low, and some interesting spots

may be unintentionally omitted. Especially, the images will be severely distorted when physicians zoom images in for detailed diagnosis. The first drawback is the nature of passive endoscopy. Some papers have presented approaches for the autonomous moving function [5], [6]. Very few papers address the solutions of the second drawback. Increasing resolution may alleviate the second problem; however, it would result in significant power consumption in RF transmitter. Hence, applying image compression is necessary for saving the power dissipation of RF transmitter. The paper [7] provides a thor-ough review on GI image compression and motivated our re-search. To overcome the second drawback, our previous work [8] has been successfully attached an ultra-low-power image compressor to the CMOS sensor to deliver a compressed 512-by-512 image while the RF transmission rate is at 2 megabits per second. In applications of capsule endoscopy, it is im-perative to consider battery life/performance tradeoffs. Instead of applying state-of-the-art video compression techniques, we proposed a simplified image compression algorithm, called GICam, in which the memory size and computational load can be significantly reduced. The experimental results shows that the GICam image compressor only costs 31 K gates at 2 frames per second, consumes 14.92 mW, and reduces the image size by 75% at least. Although the GICam image processor can successfully solve the second drawback, we expect to further reduce the power dissipation of GICam image processor itself in order to extend the battery life of capsule endoscope. Therefore, in this paper, we proposed a subsample-based GICam image compressor, called SGICam, to reduce the power dissipation of compression process. The SGICam mainly uses the subsample technique to pick out the necessary image pixels into the compression process for G1, G2 and B signals respectively. Comparing with the GICam image processor, the SGICam image compressor can significantly save 38.5% power dissipation than GICam image one.

The rest of the paper is organized as follows. In section II, we briefly review the previous work about the GICam algorithm. Section III illustrates the analysis result of discrete cosine transform for gastrointestinal image. Section IV de-scribes the SGICam algorithm in detail. Section V shows the experimental performance of the proposed algorithm. Finally, Section VI concludes our contribution and merits of this work.

II. THE REVIEW OF GICAM IMAGE COMPRESSION ALGORITHM

Fig.1 illustrates the system diagram of the proposed capsule endoscope. We attached an ultra-low-power image compressor to the CMOS sensor to deliver a compressed 512-by-512 image while the RF transmission rate is at 2 megabits per second. To reduce the buffer size between the CMOS sensor and the image compressor, the scanline controller is dedicated to scan out R, G1, G2, and B signals in a certain order. Traditional image compression algorithms use the optimized quantization for YCbCr image to reduce compressed image size while the visual distortion is low. In order to quan-tize YCbCr image, the typical image compression requires two preprocessing steps that are demosaicking and the color space transformation. However, the demosaicking step requires weighted sums for color interpolation and the color space transformation requires calculation of inner products. From the view point of GICam, it is not worth it to dissipate power for both preprocessing steps as long as the compression quality and ratio are acceptable. Hence, Fig.2 illustrates the power saving on the proposed image compression. First of all, the GICam image compression directly processes raw images without demosaicking and color space transform. For a 512×512 image, the incoming image size to the 2D-DCT is 256×256×8×4 bits, where each pixel is an 8-bit datum and each of R, G1, G2, and B components has 256×256 pixels. Since the image size after preprocessing in the traditional algorithm is 512×512×8×3 bits, the computational load of 2D-DCT and quantization is reduced by the factor of 3.

3

2

1

4

5

6

7

8

9

Fig. 1. The system structure of GICam. (1: Len; 2,3: LEDs; 4: CMOS sensor; 5: Image compressor; 6: Scanline controller; 7: Battery; 8: RF transmitter; 9: Antenna.

Traditional compression algorithms employ the YCbCr quantization to earn a good compression ratio while the visual distortion is minimized, based on the factors related to the sensitivity of the human visual system (HVS). However, for the sake of power saving, our compression rather uses the RGB quantization [9] to save the computation of demosaicking and color space transformation. As mentioned above, the advantage of applying RGB quantization is two-fold: saving the power dissipation on preprocessing steps and reducing the computing load of 2D-DCT and quantization. Moreover, to reduce the hardware cost and quantization power dissipation, we modified the RGB quantization tables and the quantization multipliers are power-oftwo’s. In GICam, the Lempel-Ziv (LZ)

coding [10] is employed for the entropy coding. The reason why we adopted the LZ coding as the entropy coding is that the LZ encoding does not need look-up tables and complex computation. Thus, the LZ encoding can consume less power and use smaller silicon size than the other candidates, such as the Huffman encoding and the arithmetic coding. The target compression performance of the GICam image compression is to reduce image size by 75% at least. To meet the specification, given the quantization tables, we exploited the cost-optimal LZ coding parameters. There are two parameters in the LZ coding to be determined; they are the window size, w, and the maximum matching length, l. (64, 16) is the minimum (w, l) set to meet the compression ratio requirement by simulating with 12 endoscopic pictures.

When comparing the proposed image compression with the traditional one in [8], the power dissipation of GICam image compressor can save 98.2% because of the reduction of mem-ory requirement. However, extending the utilization of battery life for a capsule endoscope is still an very important issue. According to the power analysis generated by PrimePowerT M

, the total power dissipation of GICam image compressor is 14.92 mW while operating at 1.8 V, in which, the power consumption of logic part is 5.52 mW, and the memory blocks generated by Artisan memory compiler consume 9.40 mW. The memory access dissipates the most power in GICam image compression. Therefore, in order to achieve the target of expending the battery life, how to efficiently reduce the memory access is necessary.

Demosaicking 2-D DCT Color Space Transform 2-D DCT 2-D DCT Quantization Y-table Quantization Cb-table Quantization Cr-table Entropy Coding Raw Image Compressed Y image Compressed Cbimage Compressed Crimage (a) 2-D DCT 2-D DCT 2-D DCT 2-D DCT Quantization G-table Quantization R-table Quantization G-table Quantization B-table Compression Image for G1 Compression Image for R Compression Image for G2 Compression Image for B Raw Image Entropy Coding Entropy Coding Entropy Coding Entropy Coding Entropy Coding Entropy Coding (b)

Fig. 2. (a) A typical image compression algorithm. (b) The GICam image compression algorithm.

III. THE ANALYSIS OF DISCRETE COSINE TRANSFORM BASIC FUNCTIONS FOR

GASTROINTESTINAL IMAGES

According to the observation from twelve tested GI images, obviously cardinal ingredient appears on tested ones.

There-fore, the red signal plays a decisive role in the raw image. For the green signal, it plays a secondary role because it is the most component in luminance. Finally, the blue signal is very indecisive after comparing with red and green ones. To verify the validity of observation mentioned above, we first use the 8-by-8 discrete cosine transform (DCT) to transfer the spatial domain into the frequency domain for each of R, G1, G2 and B components. Next, we calculates the average values of all direct current (DC) coefficients for each of R, G1, G2 and B components and the results are shown in Fig.3 (a). In Fig.3 (a), the R component has the most average value than G1, G2 and B components for each tested images and hence it can be clearly explain that why obviously cardinal ingredient appears on tested images. Except the analysis of DC domain, we also calculates the average variance values of all alternating current (AC) coefficients in order to particularity observe the variation of energy for each of R, G1, G2 and B components and the calculation results are shown in Fig.3 (b). In Fig.3 (b), the R component has the most horizontal, vertical and diagonal energy for each testing images, the secondary are G1 and G2 components and the last is B component.

Based on the analysis result mentioned above, the R com-ponent is very decisive for GI images and it needs to be compressed completely. However, the G1, G2 and B com-ponents do not need to be compressed completely because their importance are less than the R component. Therefore, in order to efficiently reduce the memory access to expend the battery life of capsule endoscopy, the datum of G1, G2 and B components should be appropriately decreased according to the proportion of their importance before the compression process. In this paper, we successfully propose a subsample-based GICam image compression algorithm and the proposed algorithm firstly uses the subsample technique to reduce the incoming datum of G1, G2 and B components before compression process. Next section will describe the proposed algorithm in detail.

IV. THE SUBSAMPLE-BASED GICAM IMAGE COMPRESSION ALGORITHM

Fig.4 illustrates the SGICam compression algorithm. For a 512×512 raw image, the raw image firstly divides into four parts, namely, R, G1, G2, B components and each of R, G1, G2, and B components has 256×256 pixels. For the R component, the incoming image size to the 2D-DCT is 256×256×8 bits, in which, the incoming image datum are completely compressed because of the importance itself in GI images. Except the R component, the SGICam algorithm can use the appropriate subsample ratio to pick out the necessary image pixels into the compression process for G1, G2 and B components and Eq.1 and Eq.2 are formulas for the subsample technique. Where, SM16:2m is the subsample

mask for the subsample ratio 16-to-2m as shown in Eq.1 and the subsample mask SM16:2m is generated from basic

mask as shown in Eq.2. The type of subample direction is block-based, when certain of positions in the subsample mask are one, their pixels in the same position will be

0 100 200 300 400 500 600 700 800 Test Picture ID A ve ra ge A C V ar ia nc e G1 G2 R B 1 2 3 4 5 6 7 8 9 10 11 12 ˃ ˅˃˃ ˇ˃˃ ˉ˃˃ ˋ˃˃ ˄˃˃˃ ˄˅˃˃ ˄ˇ˃˃ ˧˸̆̇ʳˣ˼˶̇̈̅˸ʳ˜˗ ˔̉˸̅˴˺˸ʳ˗˖ ˚˄ ˚˅ ˥ ˕ 0 200 400 600 800 1000 1200 1400 Test Picture ID A ve ra ge D C G1 G2 R B 1 2 3 4 5 6 7 8 9 10 11 12 (a) (b)

Fig. 3. (a) The avegage values of all DC coefficients for 12 testing GI images. (b) The average variance values of all AC coefficients for 12 testing GI images. Raw Image R G1 G2 B 2:1 Subsample 2:1 Subsample 4:1 Subsample 2-D 4-by-4 DCT Entropy Coding Compression Image For G1 2-D 4-by-4 DCT Entropy Coding Compression Image For G2 Non-compression Image For B 2-D 8-by-8 DCT Entropy Coding Compression Image For R 4-by-4 Quantization G-table 4-by-4 Quantization G-table Quantization R-table 4-by-4 Zig-Zag Scan 4-by-4 Zig-Zag Scan 8-by-8 Zig-Zag Scan

Fig. 4. The SGICam image compression algorithm.

compressed, otherwise they are not processed. For the G1 and G2 components, the low subsample ratio needs to be assigned because of considering their secondary importance in GI images. Thus, the 2:1 subsample ratio is the candidate one and the subsample pattern is shown in Fig.5 (a). Finally, for the B component, the 4:1 subsample ratio is assigned and the subsample pattern is shown in Fig.5 (b). In the SGICam image compression algorithm, the 8-by-8 2D-DCT is still used to transfer the R component. However, the 4-by-4 2D-DCT is used for G1 and G2 components because the incoming datum are reduced by subsample technique. Moreover, the G quantization table is also further modified and shown in the Fig.6. Finally, the B component is directly transmitted; not be compressed, after extremely decreasing the incoming datum. Due to non-compression for the B component, the 8-by-8 and 4-by-4 Zig-Zag scanning techniques are added into the SGICam to further increase the compression rate for R, G1 and G2 components before entering the entropy encoding. In the SGICam, the Lempel-Ziv (LZ) coding [10] is also employed

for the entropy coding because of non-look-up tables and low complex computation.

SM16:2m(i, j) = BM16:2m(i mod 4, j mod 4)

m= 1, 2, 3, 4, 5, 6, 7, 8. (1) BM16:2m=     u(m − 1) u(m − 5) u(m − 2) u (m − 6) u(m − 7) u(m − 3) u(m − 8) u (m − 4) u(m − 2) u(m − 5) u(m − 1) u (m − 6) u(m − 7) u(m − 3) u(m − 8) u (m − 4)    

where u(n) is a step function,u (n) = ½

1, f or n ≥ 0 0, f or n < 0.

(2)

v

Fig. 5. (1) 2:1 subsample pattern. (2) 4:1 subsample pattern.

Fig. 6. The modified G quantization table.

V. EXPERIMENTAL RESULTS

In section IV, we have been particularly introduced how to efficiently decrease the incoming datum with the subsmaple technique in the SGICam compression algorithm and then the SGICam compressor will be experimentally analyzed the performance about the compression rate, the quality degrada-tion and the ability of power saving. First of all, the target compression performance of the SGICam image compres-sion is to reduce image size by 75% at least. To meet the specification, we have to exploit the cost-optimal LZ coding parameters. There are two parameters in the LZ coding to be determined; they are the window size, w, and the maximum matching length,l. The larger the parameters, the higher the compression ratio but the higher the implementation cost. In addition, there are two kinds of LZ codings in the SGICam compressor, one is R(w, l) for R component and the other is

G(w, l) for G1 and G2 components. As per the experimental results shown in the Fig.7, we set the values of parameters by using the compression ratio of 4:1 as the threshold. Our goal is to determine the minimum R(w, l) and G(w, l) sets under the constraint of 4:1 compression ratio. The results in Fig.7 are collected by simulating the behavior model of SGICam compressor; it is generated by MATLAB. As seen in Fig.7, simulating with 12 endoscopic pictures, (32, 64) and (16,16) are the minimum R(w, l) and G(w, l) sets to meet the compression ratio requirement. Using (32, 64) and (16,16) as 5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O  5:O *:O 

Fig. 7. The simulation results of the SGICam image compression.

the parameter sets, in Table I, we can see the performance in terms of the quality degradation and compression ratio. The result shows that the degradation of decompressed images is quite low while the average PSNR is 32.18 dB. The original image involved in the PSNR calculation is the Bayer pattern image. According to the objective criterion of medical doctors, the PSNR higher than 30 dB is acceptable. To demonstrate the results, Fig.8 illustrates the compression quality of two test pictures. The difference between the original image and the decompressed image is invisible. To validate the SGICam

TABLE I

THE SIMULATION RESULTS OF TWELVE TESTED PICTURES.

Test Picture ID PSNR Compression rate

(dB) (%) 1 31.5 78.42 2 34.78 80.85 3 30.57 77.85 4 30.48 76.55 5 33.68 82.51 6 33.26 81.12 7 31.78 80.32 8 34.84 83.27 9 31.07 78.42 10 31.5 78.52 11 30.54 77.76 12 32.16 79.81 Average 32.18 79.62

image processor, we used the FPGA board of Altera APEX 2100 K to verify the function of the SGICam image processor. After FPGA verification, we used the TSMC 0.18 µm 1P6M process to implement the SGICam image compressor. When operating at 1.8 V, the power consumption of logic part is 3.88

(a) (b)

Fig. 8. (a) Demosaicked image from the fourth raw image. (b) Demosaicked image from the fourth decompressed image.

mW, estimated by using PrimePowerT M

. The memory blocks are generated by Artisan memory compiler and consume 5.29 mW. The total power consumption is 9.17 mW for the proposed SGICam image compressor. When comparing the proposed SGICam image compressor with our previous GICam one in Table II, the power dissipation can further save 38.5% under the approximate condition of quality degradation and compression ratio because of the reduction of memory requirement for G1, G2 and B components.

TABLE II

THE COMPARISON OF PROPOSED IMAGE COMPRESSION ANDGICAM IMAGE COMPRESSION.

GICam SGICam

image compressor [8] image compressor

Average PSNR 32.51 dB 32.18 dB (-0.33) Average compression 79.65 % 79.62 % (-0.03%) rate Average power 14.92 mW 9.17 mW (38.53%) dissipation VI. CONCLUSION

This paper presents an imrpoved ultra-low-power subsample-based GICam image compression processor for capsule endoscope or swallowable imaging capsules. In order to further extend the battery life of capsule endoscope, we firstly make use of the subsample technique to reduce the memory requirements of G1, G2 and B components according to the analysis results of DC/AC coefficients in 2D-DCT domain. As shown in the simulation result, the proposed image compressor can efficiently save the power dissipation than GICam one by 38.5% and reduces the video size by 75 percents at least.

REFERENCES

[1] F. Gong. P. Swain, and T. Mills, “Wireless endoscopy,” Gastrointestinal Endoscopy, vol.51, no. 6, pp. 725-729, June 2000.

[2] H. J.Park, H.W. Nam, B.S. Song, J.L. Choi, H.C. Choi, J.C. Park, M.N. Kim, J.T. Lee, and J.H. Cho, “Design of bi-directional and multi-channel miniaturized telemetry module for wireless endoscopy,” in Proc. of the 2nd Annual Intl IEEE-EMBS Special Topic Conference on Microtechnologies in Medicine and Biology, May 2-4, 2002, Madison, USA, pp. 273-276.

[3] http://www.givenimaging.com/Cultures/en-US/given/english [4] http://www.rfsystemlab.com/

[5] M. Sendoh, k. Ishiyama, and K.-I. Arai, “Fabrication of Magnetic Actuator for Use in a Capsule Endoscope,” IEEE Trans. On Magnetics, vol. 39, no. 5, pp. 3232-3234, September 2003.

[6] Louis Phee, Dino Accoto, Arianna Menciassi*, Cesare Stefanini, Maria Chiara Carrozza, and Paolo Dario, “Analysis and Development of Locomotion Devices for the Gastrointestinal Tract,” IEEE Trans. On Biomedical Engineering, vol. 49, no. 6, JUNE 2002.

[7] Kim CY, “Compression of color medical images in gastrointestinal endoscopy: a review,” Medinfo, pp. 1046-50, 1998.

[8] M. Lin, L. Dung, and P. Weng, “An Ultra Low Power Image Compressor for capsule Endoscope”, BioMedical Engineering Online 2006, vol. 5:14. [9] H.A. Peterson, H. Peng, J. H. Morgan, and W. B. Pennebaker, “Quanti-zation of color image components in the DCT domain” SPIE , Human Vision, Visual Processing, and Digital Display II, vol.1453, 1991. [10] J.Ziv and A. Lempel, “A universal algorithm for sequential data

com-pression,” IEEE Trans. On Inform. Theory, vol. 23, pp. 337-343, May 1977.