Towards robust neural models for fine-grained image recognition