Ray Space Transform Interpolation with Convolutional Autoencoder

In this paper we propose an algorithm for the reconstruction of the Ray Space Transform (RST) through the use of neural networks. In particular, our aim is to reconstruct the magnitude of the RST acquired from a linear microphone array, as if the array were composed by a larger amount of microphones. This is useful for applications that need a higher RST resolution when only a limited amount of microphones can be used due to practical constraints or physical limitations. The proposed solution leverages recent advancements in deep learning as it is based on a fully convolutional autoencoder. To validate our method, we show through a simulative campaign that it is possible to improve sound source localization using the reconstructed RST compared to the use of the original RST.