MegaLight

Abstract

Screen-camera communication using dynamic barcode streaming has emerged as a convenient and secure method for short-range, impromptu device-to-device communication. Conventional dynamic barcode systems adopt a rule-based approach to recognize the color barcode stream in the receiver, which is empirical, inflexible, and lacks self-adaptiveness. In this paper, we propose a novel solution framework for color barcode stream recognition basing on machine learning techniques. By including a number of training frames into the barcode stream to build a classification model, the proposed framework can achieve high accuracy in color barcode recognition and is adaptive to different ambient lighting conditions (without sudden changes during transmission). A semi-supervised learning approach basing on the Mixture of Experts (MoE) model is further proposed to reduce the start-up time. We implement MegaLight on both black-white and color barcode systems. Extensive experiments demonstrate that MegaLight can significantly reduce the frame demodulation error and reach up to 3x improvement in system goodput comparing to conventional barcode stream recognition approaches.