Learning Deep Bilinear Transformation for Fine-grained Image Representation