TR2019-132

DNN-based Simultaneous Screen-to-Camera and Screen-to-Eye Communications



Simultaneous screen-to-camera and screen-to-eye communications, i.e., watermarking, have been proposed in visible light communications. The main purpose of such communications is to provide many data bits for camera devices and visual information for human eyes by using a common displayed image. To this end, the existing studies leverage the capability discrepancy and distinctive features between the human vision system and camera devices. However, the existing techniques mainly require high refresh rates in both screen and camera devices to achieve better throughput while keeping high visual quality. In this paper, we propose a novel transmission scheme for efficient simultaneous screen-to-camera and screento-eye communications without a need of high refresh rates. Specifically, we use deep convolutional neural networks (DCNN)- based watermark encoder and decoder to embed many bits into high-quality images, and then to maximize throughput from the bit-embedded image. With end-to-end adversarial learning, the encoder networks learn a mapping function to embed digital data into an original image based on a perceptual loss function while the decoder networks also learn a mapping function from the bitembedded image to the data bits based on a cross-entropy loss function. From the evaluations, we show that the proposed watermark encoding and decoding networks yield high throughput from the bit-embedded images compared with a simple DCNNbased watermarking. In addition, the bit-embedded images on the screen achieve high quality for human perception.