基于融合池化和注意力增强的细粒度视觉分类网络  

Fine-Grained Visual Classification Network Based on Fusion Pooling and Attention Enhancement

作　　者：肖斌[1] 郭经伟 张兴鹏 汪敏 XIAO Bin;GUO Jingwei;ZHANG Xingpeng;WANG Min(School of Computer Science,Southwest Petroleum University,Chengdu 610500;School of Electrical Engineering and Information,Southwest Petroleum University,Chengdu 610500)

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]西南石油大学电气信息学院,成都610500

出　　处：《模式识别与人工智能》2023年第7期661-670,共10页Pattern Recognition and Artificial Intelligence

基　　金：四川省科技创新人才基金项目(No.2022JDRC0009);西南石油大学自然科学“启航计划”项目(No.2022QHZ023)资助。

摘　　要：细粒度视觉分类核心是提取图像判别式特征.目前大多数方法引入注意力机制,使网络聚焦于目标物体的重要区域.然而,这种方法只定位到目标物体的显著特征,无法囊括全部判别式特征,容易混淆具有相似特征的不同类别.因此,文中提出基于融合池化和注意力增强的细粒度视觉分类网络,旨在获得全面判别式特征.在网络末端,设计融合池化模块,包括全局平均池化、全局top-k池化和两者融合的三分支结构,获得多尺度判别式特征.此外,提出注意力增强模块,在注意力图的引导下通过注意力网格混合模块和注意力裁剪模块,获得2幅更具判别性的图像参与网络训练.在细粒度图像数据集CUB-200-2011、Stanford Cars、FGVC-Aircraft上的实验表明文中网络准确率较高,具有较强的竞争力.The core of fine-grained visual classification is to extract image discriminative features.In most of the existing methods,attention mechanisms are introduced to focus the network on important regions of the object.However,this kind of approaches can only locate the salient feature and cannot cover all discriminative features.Consequently,different categories with similar features are easily confusing.Therefore,a fine-grained visual classification network based on fusion pooling and attention enhancement is proposed to obtain comprehensive discriminative features.At the end of the network,a fusion pooling module is designed with a three-branch structure to obtain multi-scale discriminative features.The three-branch structure includes global average pooling,global top-k pooling and the fusion of the previous two.In addition,an attention enhancement module is proposed to gain two more discriminative images through attention grid mixing module and attention cropping module under the guidance of attention maps.Experiments on fine-grained image datasets,CUB-200-2011,Stanford Cars and FGVC-Aircraft,verify the high accuracy rate and strong competitiveness of the proposed network.

关 键 词：细粒度视觉分类 融合池化 注意力机制 数据增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]