CNN- Based Demodulation of Color Shift Keying in Screen Camera Communications

Besides displaying images and text for computer user interface, screens can be used to transmit information to another computer by exhibiting an array of colors that can be acquired by a digital camera and demodulated to recover the broadcasted information. This kind of modulation is known as color shift keying (CSK) and can be arranged into a M-ary communications system. In this paper, we report the results of demodulating color shift keying symbols employing convolutional neural networks (CNN) targeting a complex scenario where the received image is an arrange of
$128 \times 128$
symbols featuring a 4-CSK communications system, which comprises the aqua, fuchsia, yellow and black colors. Models are trained, validated and evaluated for 1, 9, 25 and 49 pixels per cell and compared with a maximum likelihood demodulator. Our symbol error rate results show that the proposed CNN demodulator exceed the maximum likelihood demodulator and the larger the pixels per cell involved in the demodulation process the larger the benefit from the CNN demodulator. In that sense, a symbol error rate reduction equivalent to 1478.8 times compared to the one of the maximum likelihood estimator is obtained when 49 pixel are employed for demodulation under our environment settings. This comes at the cost of high demands in terms of hardware, data handling, processing techniques and time spent in the procedures excluding the model application.