Attention-Guided CutMix Data Augmentation Network for Fine-Grained Bird Recognition

Large inter-class variations, intra-class similarities and the difficulties in collectiing training samples are three major challenges in fine-grained bird recognition. Hence, it's essential for accurate classification that achieving a discriminative feature representation from birds’ parts. In this work, we propose Attention-Guided CutMix Data Augmentation Network (AGCN) to exercise the network to pay more attention on subtle features about birds’ parts. Firstly, we generate feature maps and attention maps from an image by a backbone network. Specially, each attention map contains discriminative information of a bird's part. Then local features can be further extracted by the element-wise multiplication between feature maps and attention maps. Next, to reduce overfitting and optimize performance, we design a data augmentation strategy according to region-level replacement. For each training image, based on an attention map, a bird's discriminative region is located, copied and pasted into another image to generate an augmented image. In addition, a loss is designed to supervise both the learning of attention maps and the training of the network. AGCN only needs image-level category labels rather than bounding boxes/part annotations. The results state that our proposed data augmentation effectively improves the classification performance of the network and AGCN achieves excellent performance on the challenging dataset, CUB Birds.