Binaural recordings are inherently affected by head rotations. Usually, information on head rotation is not provided and remains unknown. However, knowing the head rotation can be helpful for speech enhancement or stabilizing binaural recordings. Our previous publication proposed a constraint-based optimization algorithm for blindly estimating head rotations from binaural recordings. This contribution proposes an alternative deep learning-based estimation algorithm. We evaluate both algorithms using real binaural recordings with different numbers of sound sources in a room. It turns out that our novel proposed model performs slightly better while the computation time during inference is strongly reduced.