LG-CNN: From local parts to global discrimination for fine-grained recognition

作者：

Highlights：

• This paper presents a fine-grained recognition system, without using bounding box and part information in both training and testing phase.

• We propose procedures for unsupervised part localization and global object discovery, and the localized part candidates and discovered approximate objects are taken as inputs of two-stream CNN.

• The two-stream CNN architecture is proposed to model both the Local part information and the Global discriminative information in a joint framework.

• We report superior or competitive results on public datasets CUB-200, Flower-102, Pets-37 and Caltech-101.

摘要

•This paper presents a fine-grained recognition system, without using bounding box and part information in both training and testing phase.•We propose procedures for unsupervised part localization and global object discovery, and the localized part candidates and discovered approximate objects are taken as inputs of two-stream CNN.•The two-stream CNN architecture is proposed to model both the Local part information and the Global discriminative information in a joint framework.•We report superior or competitive results on public datasets CUB-200, Flower-102, Pets-37 and Caltech-101.

论文关键词：Fine-grained recognition,Convolutional neural networks,Bilinear pooling,Local parts,Global discrimination

论文评审过程：Received 12 July 2016, Revised 17 May 2017, Accepted 1 June 2017, Available online 3 June 2017, Version of Record 9 June 2017.

论文官网地址：https://doi.org/10.1016/j.patcog.2017.06.002