弱监督对抗数据增强的细粒度视觉分类算法

Fine⁃grained visual classification algorithm enhanced by weakly supervised adversarial data

细粒度视觉分类(FGVC)要识别数百个属于同一基本类别的子类别,由于同一子类不同个体之间存在大方差和不同子类之间存在小方差,因此具有很高的挑战性。文中在注意力模型的基础上,提出了弱监督对抗数据增强网络。对于每张训练图像,通过弱监督学习生成注意力图,以表示对象的区别性部分。通过采用生成对抗网络(GAN)进行数据扩充来增加训练数据量,防止过度拟合。并且在数据增强网络的辅助下,主干网络可以更多地挖掘目标的区别性特征。实验结果表明,文中方法在通用的细粒度识别数据集上表现良好,比如今最先进的细粒度图像分类算法的准确度平均提高了3%。

Fine⁃Grained Visual Classification(FGVC)is to identify hundreds of sub⁃categories that belong to the same basic category.Because of the large variance between different individuals of the same sub⁃category and the small variance between different sub⁃categories,it is very challenging.A weakly supervised adversarial data enhancement network based on the attention model is proposed.For each training image,an attention map through weakly supervised learning to represent the distinctive part of the object is generated.By using Generative Adversarial Network(GAN)for data expansion,the amount of training data is increased to prevent overfitting.And with the assistance of the data augmentation network,the backbone network can mine more distinctive features of the target.Experimental results show that the method in this paper performs well on the general⁃purpose fine⁃grained identification data set,which is an average improvement of 3%compared to the most advanced fine⁃grained image classification algorithms.