Open AccessFine-Grained Image Classification based on Self-attention Feature Fusion and Graph-Propagation
Author(s) -
Pan Chen,
Wei Wu
Publication year - 2022
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2246/1/012067
Subject(s) - computer science , artificial intelligence , graph , salient , feature (linguistics) , class (philosophy) , benchmark (surveying) , relevance (law) , pattern recognition (psychology) , granularity , machine learning , image (mathematics) , data mining , theoretical computer science , geography , philosophy , linguistics , geodesy , political science , law , operating system
The goal of FGVC is to distinguish different subclasses under the same common class. Because data sets often have large intra-class differences and large inter-class similarity, FGVC is more challenging than traditional image classification. Previous work focused on mining salient regions in images and directly utilizing their features, but ignored that the features of regions are semantically related and the regional groups have stronger differences. In order to solve the above problem, we propose a FGVC method based on self-attention feature fusion and graph-propagation, consists of two branches, one of the branch based on the characteristics of map mining granularity characteristics of image, another branch study the inner semantic relevance between the regional characteristic vecto named CFSM, by iteration to enhance information elements and suppress useless element to improve the ability to recognize. A large number of experiments have proved the effectiveness of the proposed network, which has good accuracy and efficiency on the three benchmark data sets of Stanford Cars, FGVC-Aircraft and CUB200.