On-demand training of deep learning equalizers for rolling shutter optical camera communications

Abstract

The camera's exposure time restricts the reception bandwidth in rolling shutter-based optical camera communication links. Short exposures are preferable for communications, but under these conditions, the camera produces dark images with impracticable light conditions for human or machine-supervised applications. Alternatively, deep learning equalization stages can mitigate the effects of increasing the exposure time. These equalizers are trained using synthetic images based on the camera's exposure time and row sampling frequency. If these parameters are unknown in advance, another artificial network is used to estimate them directly for the captured images, the estimator. This estimator is trained offline using a vast number (thousands) of representative cases. This work proposes to transfer the attained knowledge from the offline pretrained estimator to the equalizer by using transfer learning techniques. In this way, the equalizers' training time is significantly reduced (435 times compared to full training). Consequently, transfer learning enables equalizers' online and on-demand training at reception without interfering with the communications. Results reveal that the complete training requires using exclusively 250 synthetic images to guarantee a communication performance with a bit error rate below 10(-4) after the equalization.