Mutual attention diversity feature fusion network-relevant fine-grained classification

目的 基于Transformer架构的网络在图像分类中表现出优异的性能。然而，注意力机制往往只关注图像中的显著性特征，而忽略了其他区域的次级显著信息，基于自注意力机制的Transformer也是如此。为了获取更多的有效信息，从有区别的潜在性特征中学习到更多的可判别特征，提出了一种互补注意多样性特征融合网络（complementary attention diversity feature fusion network，CADF），通过关注次显特征和对通道与空间特征协同编码，以增强特征多样性的注意感知。方法 CADF由潜在性特征模块（potential module，PFM）和多样性特征融合模块（diversity module，DFFM）组成。PFM模块通过聚合空间与通道中感兴趣区域得到显著性特征，再对特征的显著性进行抑制，以强制网络挖掘潜在性特征，从而增强网络对微小判别特征的感知。DFFM模块探索特征间的相关性，对不同尺寸的特征交互建模，以得到更加丰富的互补信息，从而产生更强的细粒度特征。结果 本文方法可以端到端地进行训练，不需要边界框和多阶段训练。在CUB-200-2011（Caltech-UCSD Birds-200-2011）、Stanford Dogs、Stanford Cars以及FGVC-Aircraft （fine-grained visual classification of aircraft）4个基准数据集上验证所提方法，准确率分别达到了92.6%、94.5%、95.3%和93.5%。实验结果表明，本文方法的性能优于当前主流方法，并在多个数据集中表现出良好的性能。在消融研究中，验证了模型中各个模块的有效性。结论 本文方法具有显著性能，通过注意互补有效提升了特征的多样性，以此尽可能地获取丰富的判别特征，使分类的结果更加精准。;Objective Fine-grained requirement is focused on images segmentation for such domain like multiple wild birds or vehicles-between features extraction in related to transferring benched category into more detailed subcategories. Due the subtle inter-category differences and large intra-category existed，it challenging capture specific regions-targeted classification. The mechanism are still used pay salient picture only although Transformer architecture-based network has its potentials image classification，and most latent ignored self-attention mechanism-based required be involved as well. To get effective information，discriminative features-derived representations learnt fine-grained feature，we develop a complementary fusion（CADF）network， which can extract multi-scale models from channel spatial interactions images. Method A mutual facilitated it consists two modules：1）potential module （PFM）：it different parts，and enhanced with preservation features. 2）Diversity module（DFFM）：multiple features-between information interaction modeling enhance rich feature，and parts terms module. scalable realize mutual-benefited information，and robustness discriminative. Our proposed configured PyTorch an NVIDIA 2080Ti GPU. weight parameters model initialized using ImageNet dataset-related SwinTransformer pre-trained. optimization performed AdamW optimizer momentum 0. 9 cosine annealing scheduler. batch size set 6，the learning rate backbone layer 000 1，the newly added 01，and decay 05 For training，the input resized 550×550 pixels cropped 448×448 randomly，and random horizontal flips data augmentation further. testing，the center. hyper-parameters λ = 1，β 5 Result verify effectiveness，the experiments carried out four datasets：CUB-Birds，Stanford Dogs，Stanford Cars，and FGVC-Aircraft. Each accuracy reached following percentages：92. 6%，94. 5%，95. 3% 93. 5%. ablation experiments，the effectiveness PFM DFFM verified Compared framework，it improve greatly via adding only. Swin-B + by 1. 4%，1. 4% 8% datasets CUB-Birds，Stanford Dogs Stanford Cars datasets. module-added only，the each exchange （Swin-B DFFM） also improved 4%，0. 5% It shows that CADF strong ability certain extent，and structure dataset potentially. visualization conducted regional intuitively. In study，the this Conclusion resolve problem insufficient mechanismbased extraction，we method