Unsupervised Sparse-view Backprojection via Convolutional and Spatial Transformer Networks