Adaptive Destruction Learning for Fine-grained Visual Classification

Fine-grained visual classification severs as an important task in artificial intelligence and greatly improves our daily life. However, this task is very challenging due to the difficulties in learning robust local features for categories with subtle visual differences. To address this issue, destruction learning, which focuses on destroying image global information and then reconstructing the entire objects, has been increasingly studied owing to its implicit local feature learning manner. Nevertheless, existing destruction learning approaches utilize a random image patch shuffling scheme to destroy the global information and easily bring in undesirable shuffling results. In this paper, we devise a novel adaptive destruction learning approach for fine-grained visual classification. Our method can actively learn the optimal shuffling matrix for each specific image, while simultaneously learning the optimal deep visual network. Both the shuffling matrix and deep visual network are trained in an end-to-end manner. The experimental results demonstrate that the proposed method achieves competitive state-of-the-art accuracy, with 89.8%, 95.3%, 93.6%, and 86.6% top-1 classification accuracy on CUB-200-2011, Stanford-Cars, FGVC-Aircraft, and Stanford-Dog, respectively.