This paper proposes a new tensor completion method based on tensor trains and system identification. The to be completed tensor is modeled as a low-rank tensor train system and the coordinates and corresponding tensor entries are interpreted as system inputs and outputs, respectively. A novel tensor train initialization procedure is proposed specifically for image and video completion, which ensures faster convergence of the completion algorithm. The tensor train framework is also shown to easily accommodate Total Variation and Tikhonov regularization due to their low-rank tensor train representations. Image and video inpainting experiments verify the superiority of the proposed scheme in terms of both speed, accuracy, and scalability, where a speedup of up to 60X is observed compared to state-of-the-art tensor completion methods at a similar accuracy.