Improving Generalization of Deep Networks for Inverse Reconstruction of Image Sequences