Mask-CNN: Localizing Parts and Selecting Descriptors for Fine-Grained Image Recognition