Convolutional neural network-based multimodal image fusion via similarity learning in the shearlet domain