基于双注意力核化双线性网络的细粒度图像分类

Fine-grained image classification based on dual attention kernelized bilinear CNN

针对双线性卷积网络忽略特征图中不同通道和空间位置对分类的不同作用问题,提出一种基于双注意力机制的核化双线性卷积网络模型。从通道和空间两个维度上对局部区域进行双注意力建模,通道注意力机制对通道加权,空间注意力机制对位置加权,将两个机制的注意力特征图矩阵相加后进行外积聚合。采用sigmoid核函数对外积矩阵进行核化,建模通道间的非线性关系。实验在CUB-200-2011、FGVC-Aircraft以及Standford-Cars这3个细粒度数据集上对该方法进行测试,实验结果表明,该方法在3个数据集上均优于同类方法。

Aiming at the problems that Bilinear CNN(B-CNN)ignores the different roles of each channel and spatial position in classification,a kernelized bilinear convolutional network model based on dual attention mechanism was proposed.Dual-attention modeling for the local parts was carried out from two dimensions of channel and space,in which the channel attention mechanism weighted the channel and the spatial attention mechanism weighted the position.The attention feature maps of two mechanisms were added together to perform outer product aggregation.The sigmoid kernel function was used to kernel the outer product matrix to capture the nonlinear correlation between channels.This method was evaluated on three fine-grained datasets of CUB-200-2011,FGVC-Aircraft and Stanford-Cars.Experimental results show that the method is superior to its counterparts on all three datasets.