细粒度图像之间具有高度相似的外观,其差异往往体现在局部区域,提取具有判别性的局部特征成为影响细粒度分类性能的关键。引入注意力机制的方法是解决上述问题的常见策略,为此,在双线性卷积神经网络模型的基础上,提出一种改进的双线性残差注意力网络:将原模型的特征函数替换为特征提取能力更强的深度残差网络,并在残差单元之间分别添加通道注意力和空间注意力模块,以获取不同维度、更为丰富的注意力特征。在3个细粒度图像数据集CUB-200-2011、Stanford Dogs和Stanford Cars上进行消融和对比实验,改进后模型的分类准确率分别达到87.2%、89.2%和92.5%。实验结果表明,相较原模型及其他多个主流细粒度分类算法,本文方法能取得更好的分类结果。

Fine-grained images have a highly similar appearance, and the differences are often reflected in local regions. Extracting discriminative local features plays a key role in fine-grained classification. Attention mechanism is a common strategy to solve the problems above. Therefore, we propose an improved bilinear residual attention network based on bilinear convolutional neural network model in this paper: the feature function of the original model is replaced by deep residual network with a stronger feature extraction capability, then channel attention module and spatial attention module are added between the residual units respectively to obtain different dimensions and richer attention features. Ablation and contrast experiments were performed on three fine-grained image datasets CUB-200-2011, Stanford Dogs, and Stanford Cars, the classification accuracy of the improved model reached 87.2%, 89.2% and 92.5%, respectively. Experimental results show that our method can achieve better classification results than the original model and other mainstream fine-grained classification algorithms.