基于对抗擦除的细粒度图像数据增强方法

Method of Fine-Grained Visual Data Augmentation Based on Adversarial Erasing

不同于常规图像识别,细粒度图像识别中不同类别的视觉差异往往只取决对象的细微部件,因此对于细粒度图像识别任务而言,发现对象具有判别力的细微部件发挥了重要作用。为此,文中提出了基于对抗擦除的数据增强方法(AEDA),在训练阶段首先通过特征图定位对象最具有判别力的细微部件作为增强的部件图像,然后擦除对象最具有判别力的细微部件作为增强的互补图像。通过输入部件图像可以使网络学习对象最具有判别力的细微部件,通过输入互补图像可以迫使网络发现对象其他具有判别力的细微部件。在细粒度图像识别领域的三个经典公开数据集上的实验结果表明文中所提数据增强方法可以大幅提升模型性能且全面优于基于擦除的经典数据增强方法Cutout。此外引入定位模块(AOLM),使文中所提方法识别性能进一步提升,在CUB、Aircraft、Car数据集上分别达到了88.7%、94.2%、95.3%的识别精度。同时该方法还大幅度提升了定位性能,表明其在弱监督目标定位视觉任务的潜力。

Different from conventional image recognition,the visual differences of different categories in fine-grained image recognition usually only depend on the subtle parts of the object.Therefore,it is extremely important to find out the discriminative subtle parts of the object for fine-grained image recognition.In this paper,a data augmentation based on adversarial erasing(AE-DA)is proposed.In the training stage,the most discriminative subtle part of the object is firstly located as the augmented part image through the feature map,then the most discriminative part of the object is erased as the augmented complementary image.The net-work can learn the most discriminative subtle part of the object by part image,and find out other discriminative subtle parts of the object by complementary image.The Experiments show that the proposed data augmentation can greatly improve the model perfor-mance and outperforms than the classical method Cutout based on erasing.Furthermore,the proposed method is improved and achieving the classification accuracy of 88.7%,94.2%and 95.3%on CUB,Aircraft and Cars by incorporating the location module(AOLM).At the same time,it also greatly improves the location performance,demonstrating its potential in the weakly supervised object localization task.