Online image augmentation via regional cross-attention

Aiming at overcoming undesirable fine-grained image augmentation caused by inaccurate salient region location in challenging scenarios, we propose a cross-region attention-based automatic image augmentation method. Unlike existing offline methods that randomly choose the sub-regions for image mixing, we first propose a cross-region attention mechanism, which can simultaneously explore the intro and inter-image block relation through the collaboration of fix and shift window-based cross-attention. Secondly, an image mixing sub-network is designed based on the cross-region attention. This sub-network can be easily employed in various fine-grained classification networks. After integrating the image mixing sub-network and finegrained classification network into a union, we focus on solving the two-layer optimization problem caused by the inconsistency between heterogeneous network structures. The superiority of the proposed method for fine-grained image enhancement is verified through extensive experimental evaluations on six fine-grained datasets.