结合注意力与双线性网络的细粒度图像分类

Combines Attention with Bilinear Networks for Fine-grained Image Classification

如何对识别物体进行精确定位并提取更具有表达力的特征,是细粒度图像分类算法的核心问题之一.为此,本文提出了一种基于注意力机制的双线性卷积神经网络细粒度图像分类算法(BAM B-CNN),主要工作如下:1)通过VGG-16网络获得原始图像的激活映射图,选取大于平均值的最大联通区域作为物体图像;2)使用区域建议网络(RPN)提取候选区域,结合部件注意力模型将候选区域分为k组,以各组评分最高的候选区域作为部件图像;3)在双线性网络中引入通道注意力模块,学习通道间的非线性关系,提高关键特征的表达力;4)使用分类模型结合不同层次特征的优点,提高分类精度.理论分析和试验验证均验证了所提算法的有效性.

How to locate the recognized object and extract more expressive features is one of the core problems of fine-grained image classification algorithm.To this end,this paper proposed a fine-grained image classification method named as Based on Attention Mechanism in Bilinear CNN(BAM B-CNN).The main contributions of this paper are as follows:1)The activation map of the original image is obtained through the VGG-16 network,and the maximum connecting region larger than the average value of the activation map is selected as the object image;2)Region Propose Network(RPN)was used to provide candidate regions.The candidate regions were divided into k groups by Part-Level Attention Model,and the candidate regions with the highest scores in each group were taken as part images;3)The Channel Attention Module is added to bilinear CNN to learn the nonlinear relationship among channels and improve the expression of key features;4)The classification model combines the advantages of features in different levels to improve the classification accuracy.Theoretical analysis and experiments prove the effectiveness of the proposed method in this paper.