Radio DIP - Completing Radio Maps using Deep Image Prior

Ray tracing is one of the de-facto standard method-ologies for radio channel modelling, given the geographical map of the layout. However, the channel generated by ray-tracing cannot be adapted to incorporate knowledge from real-world channel measurements. Several recent papers have proposed training a deep neural network (DNN) to compute the radio map for a given input layout. Such techniques typically require a large number of measurements, transmitters and receivers to generate the dataset needed for training the DNN, and hence can only be trained on simulated data from ray tracing. We propose an extension to these techniques, whereby we first train our DNN on simulated data, and then use a small number of measurements from a given setting to predict the path loss at all locations of interest, borrowing from a generative modelling technique called Deep Image Prior. Our simulations show that Radio DIP can achieve a RMSE of 5 dB in predicting the path loss of 50k outdoor locations, given less than 100 measurements.