Expand this Topic clickable element to expand a topic
Skip to content
Optica Publishing Group
Journal Home About Issues in Progress Current Issue All Issues Feature Issues

Using logistic regression classification for mitigating high noise-ratio advisement light-panel in rolling-shutter based visible light communications

Yu-Cheng Chuang, Chi-Wai Chow, Yang Liu, Chien-Hung Yeh, Xin-Lan Liao, Kun-Hsien Lin, and Yi-Yuan Chen

Open Access Open Access

Abstract

We propose and experimentally demonstrated a light-panel and image sensor based visible light communication (VLC) system using machine learning (ML) algorithm. The ML algorithm is compared with the traditional demodulation scheme and the experimental results show that even at very high noise-ratio (NR) light-panel display content, the proposed ML algorithm shows significant bit error rate (BER) improvement.

© 2019 Optical Society of America under the terms of the OSA Open Access Publishing Agreement

1. Introduction

Visible light communication (VLC) is one of the promising wireless communications, and the transmitter (Tx) and receiver (Rx) in visible spectrum are mature and low cost [1–3]. Different VLC systems were demonstrated [4–10]. As there are billions of mobile-phones in the markets, using the embedded complementary-metal–oxide–semiconductor (CMOS) image sensor as VLC Rx could be interesting. However, using the CMOS image sensor for VLC detection is challenging, since the frame rate of the CMOS image sensor is limited. Besides, the frame-to-frame “blind-gap” in the CMOS image sensor makes the VLC synchronization and demodulation difficult. Recently, 150 bit/s and 75 bit/s VLC systems using red-green-blue (RGB) light emitting diodes (LEDs) and CMOS image sensor were demonstrated; but the data rates were limited by the frame-rate [11,12]. Tailor-made image senor with image sensing pixels and high-speed PD pixels has been demonstrated [13]; however, this device can be costly. Using the rolling shutter effect (RSE) of a CMOS image sensor for VLC can be promising [14,15] and it was successfully illustrated that the data rate of RSE-based VLC system can be much higher than the frame rate of the CMOS image sensor [15].

In this work, we propose and experimentally demonstrate a VLC system using light-panel Tx and mobile-phone CMOS image sensor Rx with machine learning (ML) algorithm. The ML algorithm is based on logistic regression classification. The LED in the light-panel is modulated using on-off-keying (OOK); and users receive the VLC data by using their RSE-based smart-phone CMOS image sensor. As the light-panel will display different advertisements (different contents), previously we have defined the noise-ratio (NR) [9] as the figure-of-merit to evaluate the display content grayscale value contrast with respected to all-white content. By using our developed mobile-phone application program and the proposed ML algorithm, the light-panel based VLC system performance can be significantly enhanced even at a very high NR display content.

2. Architecture and ML algorithm

Figure 1(a) shows the proof-of-concept experiment of VLC system using light-panel and smart-phone CMOS image sensor. The light-panel is connected to an arbitrary waveform generator (AWG, Tektronix AFG3252C). The light-panel (Li-Cheng Corp.) has a dimension of 878 mm x 577 mm, with the output power of 22W produced by an LED light bar (Epistar) installed at the panel side-frame. Different advertising contents can be put onto the light-panel. This light-panel based VLC system is suitable for transmitting leaflets in shops, menus in restaurants and supermarkets, location information or maps in train stations. The AWG has an analog bandwidth of 240 MHz and sampling rate of 2 GS/s. The smart-phone used as the Rx is IPhone 7 + , the CMOS camera is operated at 1080p with 30 fps. The ISO used is 400 and the shutter speed is 1/30 s. The resolution of the image frame is 1080 x 1920 pixels with 16:9 ratio. We wish our developed application program can be applicable in different brands of mobile-phones; hence we select the lower frame rate of 30 fps. During the processing, if the pixel-per-bit is not an integer, we will select the nearest integer smaller than the non-integer pixel-per-bit value to avoid the sampling induced error.

 figure: Fig. 1

Fig. 1 (a) VLC system using light-panel and smart-phone CMOS image sensor. (b) flow diagram of the ML algorithm implementation to improve the VLC performance.

Download Full Size | PDF

Figure 1(b) shows the flow diagram of the ML algorithm implementation to improve the VLC performance. The VLC data packet generated by the light-panel is captured by our developed APP. By optimizing several CMOS image sensor parameters in the APP, the VLC RSE pattern (bright and dark fringes) can be captured even at low illuminance of 188 lux. As shown in Fig. 1(b), the frame image in the raw movie file will be read-in by the program. First, the location of the light-panel in an image frame is identified. This could be implemented by using region-grow light source tracking mechanism [16]. As there is a frame-to-frame “blind-gap” during the detection of the CMOS image sensor, the VLC packet header should be located; and the payload data will be extracted between two VLC packet headers. The proposed ML algorithm is based on logistic regression classification scheme [17], in which the RGB basic color RSE patterns will be separated and multiplied by different weights with proper biases. As shown in Fig. 1(b). The payload RSE pattern will be divided into the R, G, and B basic color RSE patterns. Before applying the ML algorithm, each basic color RSE pattern will be converted to a column matrix, in which the length of the column matrix represents the number of pixel rows of the light-panel shown in the image frame. As the grayscale values of different pixels in each pixel row are different due to different displayed content, we first apply row averaging to obtain an average grayscale for each pixel row. To obtain a bipolar normalized grayscale values for the subsequence ML algorithm, the z-score normalization is applied. The z-score normalized value of each pixel row is equal to the average grayscale minus the means of all the averaged grayscale values, and the whole term is divided by the standard deviation of all the averaged grayscale values. The ML algorithm is based on logistic regression classification [17], in which the R, G, and B z-score normalized matrices will be multiplied by different weights w with proper biases b. Assume the length of the VLC payload data occupies N pixels in an image frame, this means the length of the column matrix is N. We can obtain the posterior probability Pn by using the sigmoid function σ(.) as shown in Eq. (1), where n is from 1 to N

\[{P_n} = \sigma ({z_n}), {\rm{and}} {z_n} = {\bf{wx}}_{\bf{n}}^{\bf{T}} + b = {w_r}x_r^{(n)} + {w_g}x_g^{(n)} + {w_b}x_b^{(n)} + b.\]
The weight vector and RGB vector are represented in Eq. (2),
\[{\bf{w}} = ({w_r}, {w_g}, {w_b}), {\rm{and}} {{\bf{x}}_{\bf{n}}} = (x_r^{(n)}, x_g^{(n)}, x_b^{(n)}),\]
where $\[x_r^{(n)}, x_g^{(n)}, x_b^{(n)}\]$ are the nth z-score normalized values of R, G and B respectively; and $\[{w_r}, {w_g}, {w_b}\]$ are the weights of R, G and B respectively. After this, we calculate the cross entropy error function obtained from the converted probability Pn and the target value tn as illustrated in Eq. (3)
\[E({\bf{w}}{\rm{,b) = }} - \sum\limits_{n = 1}^N {{E_n}} = - \sum\limits_{n = 1}^N {\left[ {{t_n}\ln {P_n} + (1 - {t_n})\ln (1 - {P_n}{\rm{)}}} \right]} .\]
By using gradient descent method [19] to minimize the cross entropy error function; the updated weight w and bias b can be obtained as shown in Eq. (4), where τ is number of iterations, η is the learning rate
\[{{\bf{w}}^{\tau {\rm{ + 1}}}} = {{\bf{w}}^\tau }{\rm{ - }}\eta \frac{{\partial E}}{{\partial {\bf{w}}}}, {\rm{and}} {b^{\tau {\rm{ + 1}}}} = {b^\tau }{\rm{ - }}\eta \frac{{\partial E}}{{\partial b}}.\]
In our developed application program for the mobile phone, 30 fps is used. One image frame of the 30 frames is selected for the training process. Then the captured frame is then divided into the R, G, and B basic color patterns for the training; and the training processes for the R, G, and B channels are performed separately based on the gradient descent method described in Eq. (4). Finally, the thresholding can be based on the obtained converted probability Pn, as illustrated in Eq. (5). If it is greater than or equal to 0.5, it is regarded as logic 1; otherwise, it is regarded as logic 0

\[Logic = \left\{ \begin{array}{l}1, {P_n} \ge 0.5\\{\rm{0}}, {P_n} < 0.5\end{array} \right..\]

The proposed VLC system based on ML algorithm is also compared with the previous light-panel based VLC system based on traditional demodulation scheme reported in [9]. In the traditional light-panel based VLC packet demodulation, the efficient extinction-ratio (ER) enhancement [9] and adaptive thresholding, called extreme value averaging (EVA) thresholding [18] can be applied to the column matrix to identify the logic 1 and logic 0. Figure 2 illustrates the implementation of the traditional light-panel based VLC system (top) and the proposed VLC system based on ML algorithm (bottom). Figure 2(a) shows the received RSE pattern emitted by the light-panel with a high NR of 70.21%. The RSE of CMOS image sensor detection is described in [14,15]. The CMOS image sensor does not acquire the whole image frame simultaneously, instead each pixel row is activated sequentially. When the light-panel Tx is “ON” and “OFF” faster than the frame rate of the CMOS image sensor, bright (“ON”) and dark (“OFF”) fringes are captured in each image frame. By demodulating these bright and dark fringes, the transmitted VLC data can be retrieved. The definition of NR is reported in [9], and it indicates the grayscale value contrast with respected to all-white content in the light-panel. For example, the light-panel with a all-white and all-black contents have the NR = 0% and 100% respectively. After the grayscale row averaging, the RSE pattern can be converted to the grayscale pattern as shown in Fig. 2(b), and we can observe a high grayscale fluctuation due to the dark display content of the light-panel. The high signal fluctuation making the traditional thresholding scheme used in fiber-optics communications not functioning. Because of this, ER enhancement and EVA adaptive thresholding scheme [18] has been proposed, and the bit-logic is shown in Fig. 2(c). Although the previously proposed ER enhancement and EVA thresholding is efficient to demodulate the RSE pattern previously, it is not working well at long transmission VLC distance and high NR value display contents. By comparing the bi-logic demodulated by using traditional method as shown in Fig. 2(c) with the transmitted bit-logic shown in Fig. 2(d), we can still observe high bit errors particularly in the pixel rows between 200th to 300th (indicated by red dotted circles). By applying the proposed ML logistic regression classification, the z-score normalized values of the R, G, and B patterns are shown in Figs. 2(e)-2(f) respectively. We can also observe high signal fluctuations in these patterns similar to Fig. 2(b) due to the high NR value of the light-panel display content. By implementing the ML algorithm as discussed earlier in this section, the correct bit-logic pattern shown in Fig. 2(h) can be obtained.

 figure: Fig. 2

Fig. 2 Implementation of the traditional light-panel based VLC system (top) and the proposed VLC system based on ML algorithm (bottom).

Download Full Size | PDF

3. Results and discussions

Owing to different display contents of the light-panel; the signal-to-noise ratios (SNRs) received by the image sensor are different. Figure 3 shows the different display contents of the light-panel with different NR values and at different data rates. We can observe that at higher data rate, the density of bright and dark fringes increases due to the reduction of pixel per bit observed in the image frame. The light-panel based VLC experiments are performed based on different display contents at different NR values and at different data rates. The VLC transmission distances are 150 cm for all measurements. Figures 4(a)-4(c) show the BER measurements when using traditional demodulation schemes [9] and the proposed ML algorithm at NR = 0%, 39.56% and 70.21% respectively. The performance improvement of the proposed ML scheme is not very obvious when the NR of the display contents in the light-panel is low or moderate, as shown in Figs. 4(a) and 4(b). At high NR values (i.e. the display content is “dark”), the proposed ML algorithm shows significant BER improvement and can achieve a transmission data rate of 780 bit/s satisfying the forward error correction (FEC) (BER ≦ 3.8 x 10−3) requirement while the traditional scheme cannot achieve the FEC, as shown in Fig. 4(c). Figure 5(a) shows the measured pixel per bit at different data rates. When the data rate increases, the number of pixel row required to represent one bit becomes smaller; hence increasing the difficulty in demodulating the VLC RSE pattern. In the proposed ML scheme, data rate of 1.26 kbit/s can be achieved at NR = 0% representing the pixel per bit of 10.50. At high NR value of 70.21%, it becomes more difficult to demodulate the RSE pattern; hence the data rate after applying ML is 1.02 kbit/s, representing the pixel per bit = 12.48. Figure 5(b) shows the measured illuminance at different NR values (i.e. different display content) when the distance between the light-panel and the smart-phone is 150 cm. At high NR value of 70.21%, the display content is very dark and the measured illuminance is only 188 lux. This shows that even at very low illuminance, the proposed ML algorithm can successfully demodulate the RSE pattern and achieve data rate of 1.02 kbit/s. The data rate of the VLC system can be improved by increasing the frame rate of the CMOS image sensor or the light-panel brightness. Decreasing the free-space transmission distance or using a larger light-panel can also reduce the required pixel-per-bit, which can also increase the data rate.

 figure: Fig. 3

Fig. 3 Display contents of light-panel with different NR values and at different data rates.

Download Full Size | PDF

 figure: Fig. 4

Fig. 4 Measured BER curves using traditional demodulation and the proposed ML algorithm when the (a) NR = 0%, (b) NR = 39.56% and (c) NR = 70.21% respectively.

Download Full Size | PDF

 figure: Fig. 5

Fig. 5 (a) Measured pixel per bit at different data rates, and (b) illuminance at different NRs.

Download Full Size | PDF

4. Conclusion

We proposed and demonstrated a light-panel and CMOS image sensor based VLC system using ML algorithm. The ML algorithm was based on logistic regression classification, in which the R, G, and B z-score normalized matrices were multiplied by different weights w with proper biases b. The proposed VLC system based on ML algorithm was compared with the traditional demodulation scheme. At high NR value, the proposed ML algorithm shows significantly BER improvement and can achieve a transmission data rate of 780 bit/s satisfying the FEC requirement while the traditional scheme cannot achieve the FEC when the transmission distance is 150 cm.

Funding

Ministry of Science and Technology, Taiwan, ROC (MOST-107-2221-E-009-118-MY3, MOST-106-2221-E-009-105-MY3); Aim for the Top University Plan; Ministry of Education, Taiwan.

References

1. S. J. Nawaz, S. K. Sharma, S. Wyne, M. N. Patwary, and M. Asaduzzaman, “Quantum machine learning for 6G communication networks: state-of-the-art and vision for the future,” IEEE Access 7, 46317–46350 (2019). [CrossRef]  

2. H. Haas, “Visible light communication,” Proc. OFC, 1–72 (2015).

3. C. W. Chow, C. H. Yeh, Y. Liu, and Y. F. Liu, “Digital signal processing for light emitting diode based visible light communication,” IEEE Photon. Soc. Newslett. 26, 9–13 (2012).

4. Z. Wang, C. Yu, W. D. Zhong, J. Chen, and W. Chen, “Performance of a novel LED lamp arrangement to reduce SNR fluctuation for multi-user visible light communication systems,” Opt. Express 20(4), 4564–4573 (2012). [CrossRef]   [PubMed]  

5. R. Deng, J. He, and G. K. Chang, “Enhancement of signal performance in bidirectional RSE-based CVLC system employing NPM modulation scheme,” Opt. Commun. 432, 97–100 (2019). [CrossRef]  

6. H. H. Lu, Y. P. Lin, P. Y. Wu, C. Y. Chen, M. C. Chen, and T. W. Jhang, “A multiple-input-multiple-output visible light communication system based on VCSELs and spatial light modulators,” Opt. Express 22(3), 3468–3474 (2014). [CrossRef]   [PubMed]  

7. B. Janjua, H. M. Oubei, J. R. Durán Retamal, T. K. Ng, C. T. Tsai, H. Y. Wang, Y. C. Chi, H. C. Kuo, G. R. Lin, J. H. He, and B. S. Ooi, “Going beyond 4 Gbps data rate by employing RGB laser diodes for visible light communication,” Opt. Express 23(14), 18746–18753 (2015). [CrossRef]   [PubMed]  

8. C. Lee, C. Shen, H. M. Oubei, M. Cantore, B. Janjua, T. K. Ng, R. M. Farrell, M. M. El-Desouki, J. S. Speck, S. Nakamura, B. S. Ooi, and S. P. DenBaars, “2 Gbit/s data transmission from an unfiltered laser-based phosphor-converted white lighting communication system,” Opt. Express 23(23), 29779–29787 (2015). [CrossRef]   [PubMed]  

9. C. W. Chow, R. J. Shiu, Y. C. Liu, C. H. Yeh, X. L. Liao, K. H. Lin, Y. C. Wang, and Y. Y. Chen, “Secure mobile-phone based visible light communications with different noise-ratio light-panel,” IEEE Photonics J. 10(2), 7902806 (2018). [CrossRef]  

10. C. W. Chow, R. J. Shiu, Y. C. Liu, X. L. Liao, K. H. Lin, Y. C. Wang, and Y. Y. Chen, “Using advertisement light-panel and CMOS image sensor with frequency-shift-keying for visible light communication,” Opt. Express 26(10), 12530–12535 (2018). [CrossRef]   [PubMed]  

11. P. Luo, M. Zhang, Z. Ghassemlooy, H. L. Minh, H. M. Tsai, X. Tang, L. C. Png, and D. Han, “Experimental demonstration of RGB LED-based optical camera communications,” IEEE Photonics J. 7(5), 7904242 (2015). [CrossRef]  

12. C. W. Chow, R. J. Shiu, Y. C. Liu, Y. Liu, and C. H. Yeh, “Non-flickering 100 m RGB visible light communication transmission based on a CMOS image sensor,” Opt. Express 26(6), 7079–7084 (2018). [CrossRef]   [PubMed]  

13. I. Takai, S. Ito, K. Yasutomi, K. Kagawa, M. Andoh, and S. Kawahito, “LED and CMOS image sensor based optical wireless communication system for automotive applications,” IEEE Photonics J. 5(5), 6801418 (2013). [CrossRef]  

14. C. Danakis, M. Afgani, G. Povey, I. Underwood, and H. Haas, “Using a CMOS camera sensor for visible light communication,” Proc. OWC2012, 1244–1248. [CrossRef]  

15. C. W. Chow, C. Y. Chen, and S. H. Chen, “Visible light communication using mobile-phone camera with data rate higher than frame rate,” Opt. Express 23(20), 26080–26085 (2015). [CrossRef]   [PubMed]  

16. K. Liang, C. W. Chow, and Y. Liu, “Mobile-phone based visible light communication using region-grow light source tracking for unstable light source,” Opt. Express 24(15), 17505–17510 (2016). [CrossRef]   [PubMed]  

17. C. M. Bishop, Pattern Recognition and Machine Learning (Springer, 2006).

18. C. W. Chen, C. W. Chow, Y. Liu, and C. H. Yeh, “Efficient demodulation scheme for rolling-shutter-patterning of CMOS image sensor based visible light communications,” Opt. Express 25(20), 24362–24367 (2017). [CrossRef]   [PubMed]  

Cited By

Optica participates in Crossref's Cited-By Linking service. Citing articles from Optica Publishing Group journals and other participating publishers are listed here.

Alert me when this article is cited.


Figures (5)
Fig. 1
Fig. 1 (a) VLC system using light-panel and smart-phone CMOS image sensor. (b) flow diagram of the ML algorithm implementation to improve the VLC performance.

View in Article | Download Full Size | PDF

Fig. 2
Fig. 2 Implementation of the traditional light-panel based VLC system (top) and the proposed VLC system based on ML algorithm (bottom).

View in Article | Download Full Size | PDF

Fig. 3
Fig. 3 Display contents of light-panel with different NR values and at different data rates.

View in Article | Download Full Size | PDF

Fig. 4
Fig. 4 Measured BER curves using traditional demodulation and the proposed ML algorithm when the (a) NR = 0%, (b) NR = 39.56% and (c) NR = 70.21% respectively.

View in Article | Download Full Size | PDF

Fig. 5
Fig. 5 (a) Measured pixel per bit at different data rates, and (b) illuminance at different NRs.

View in Article | Download Full Size | PDF

Equations (7)

Equations on this page are rendered with MathJax. Learn more.

P n = σ ( z n ) , a n d z n = w x n T + b = w r x r ( n ) + w g x g ( n ) + w b x b ( n ) + b .
w = ( w r , w g , w b ) , a n d x n = ( x r ( n ) , x g ( n ) , x b ( n ) ) ,
x r ( n ) , x g ( n ) , x b ( n )
w r , w g , w b
E ( w , b ) = − ∑ n = 1 N E n = − ∑ n = 1 N [ t n ln ⁡ P n + ( 1 − t n ) ln ⁡ ( 1 − P n ) ] .
w τ + 1 = w τ − η ∂ E ∂ w , a n d b τ + 1 = b τ − η ∂ E ∂ b .
L o g i c = { 1 , P n ≥ 0.5 0 , P n < 0.5 .
Select as filters
Select Topics Cancel
Top
       
© Copyright 2024 | Optica Publishing Group. All rights reserved, including rights for text and data mining and training of artificial technologies or similar technologies.
Privacy | Terms of Use