多感知兴趣区域特征融合的图像识别方法  

Image recognition method based on multi-perceptual interest region feature fusion

作　　者：闫涵 张旭秀[1] 张净丹 YAN Han;ZHANG Xuxiu;ZHANG Jingdan(School of Electrical Information Engineering,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学电气信息工程学院,辽宁大连116028

出　　处：《智能系统学报》2021年第2期263-270,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61471080/F010408);国家支撑计划(2015BAF20B02);国家留学基金委资助计划(201608210308);辽宁省自然科学基金指导计划(2019-ZD-0108).

摘　　要：针对自然图像识别过程中不同深度学习模型关注兴趣区域不同的现象,本文引入深度卷积神经网络融合机制,结合深度迁移学习方法,给出了一种基于多感知兴趣区域特征融合的图像识别方法。本文将迁移学习方法引入牛津大学视觉组网络模型(visual geometry group network,VGGNet)和残差网络模型(residual network,ResNet),通过对单个分类模型进行热力图可视化及特征可视化,得到了不同网络模型关联的特征区域不一样的结论。然后在此基础上分别设计特征拼接、特征融合加特征拼接及融合投票方法将不同模型特征进行融合,得到3种新的融合模型。实验结果表明,本文方法在Kaggle数据集上的识别准确率高于VGG-16、VGG-19、ResNet-50、DenseNet-201模型。This paper presents the deep convolution neural network fusion mechanism and proposes an image recognition method based on multi-perceptual interest region feature fusion in combination with the deep-migration learning method.This is to solve the problem of different deep-learning models used on different interest regions when they recognize a natural image.The migration learning method is applied to the convolution neural net architectures,namely VGG and ResNet networks.Then,through the visualization of the heat map and the features of single classification model,a conclusion is drawn that the characteristic regions associated with different network models are different.Based on this,the methods of feature splicing,feature fusion and splicing,and fusion voting systems are designed to fuse different model features,obtaining three new fusion models.The experimental results show that the recognition accuracy of this method on Kaggle dataset is higher than that of VGG-16,VGG-19,ResNet-50,and DenseNet-201 models.

关 键 词：深度学习 图像识别 迁移学习 特征融合 集成学习 特征提取 CAM可视化 视觉组网络模型 残差网络模型 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]