Deep-Learning-Based Adaptive Symbol Decision for Visual MIMO System with Variable Channel Modeling

A channel modeling method and deep-learning-based symbol decision method are proposed to improve the performance of a visual MIMO system for communication between a variable-color LED array and camera. Although image processing algorithms using color clustering are available to correct distorted color information in a channel, color-similarity-based approaches are limited by real-world distortions; to overcome such limitations, symbol decision is defined as a multiclass classification problem. Further, to learn a robust classifier against channel distortion, a deep neural network learning technique is applied to adaptively determine symbols from channel distortion. The network designed herein comprises the channel identification and symbol decision modules; the channel identification module extracts a channel identification vector for symbol determination from an input image using a two-dimensional deep convolutional neural network (CNN); the symbol decision module then generates a feature map by combining the channel identification vector and information on adjacent symbols to determine the symbol via learning correlations between adjacent symbols using a one-dimensional CNN. The two modules are connected together and learned simultaneously in an end-to-end manner. We also propose a new channel modeling method that intuitively reflects real-world distortion factors rather than the conventional additive white Gaussian noise channel to efficiently train deep-learning networks. Lastly, in the proposed channel distortion environment, the proposed method shows performance improvement by an average of about 41.8% (up to about 54.8%) compared to the existing Euclidean distance method, and about 6.3% (up to about 9.2%) on average compared to the SVM method.