Deep Learning Approach to Optical Camera Communication Receiver Design

This paper investigates a deep learning (DL) framework for designing optical camera communication (OCC) systems where a receiver is realized with optical cameras capturing images of transmit LEDs. The optimum decoding strategy is formulated as the maximum a posterior (MAP) estimation with a given received image. Due to the absence of analytical OCC channel models, it is challenging to derive the closed-form MAP detector. To address this issue, we employ a convolutional neural network (CNN) model at the OCC receiver. The proposed CNN approximates the optimum MAP detector that determines the most probable data symbols by observing an image of the OCC transmitter implemented by dot LED matrices. The supervised learning philosophy is adopted to train the CNN with labeled images. We collect training samples in real-measurement scenarios including heterogeneous background noise and distance setups. As a consequent, the proposed CNN-based OCC receiver can be applied to arbitrary OCC scenarios without any channel state information. The effectiveness of our model is examined in the real-world OCC setup with Raspberry Pi cameras. The experimental results demonstrate that the proposed CNN architecture performs better than other DL models.