基于注意力机制的DenseNet模型的树种识别应用  

Application of Tree Species Identification Based on DenseNet Model with Attention Mechanism

作　　者：宋宇鹏 边继龙[1] 安翔 张锡英[1] SONG Yupeng;BIAN Jilong;AN Xiang;ZHANG Xiying(School of Information and Computer Engineering,University of Northeast Forestry,Harbin 150040,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150040

出　　处：《实验室研究与探索》2020年第7期122-127,173,共7页Research and Exploration In Laboratory

基　　金：国家自然科学基金项目(No.31500587);中央高校基本科研业务费专项资金项目(No.2572016BB11);黑龙江省科学基金项目(No.F2018002)。

摘　　要：随着图像识别分类技术的发展,该技术被人们应用到工农业生产各个领域,以提高其工作质量和效率。在特殊领域背景复杂数据集分类任务中,为增强神经网络的分类能力,降低参数冗余,提高训练效率,提出一种基于注意力机制的DenseNet模型。该神经网络能够通过添加注意力机制获取图像重要信息,以解决数据敏感问题,提高网络整体性能。在复杂树种叶片公开数据集Leafsnap和公共数据集SVHN上分别取得了91.25%和98.27%的分类精确率。实验结果表明,基于注意力机制的DenseNet模型分类效果明显优于其他网络模型。With the development of image recognition and classification technologies,people gradually apply the technologies to various fields to improve their work quality and work efficiency.In the task of classifying complex data sets in special fields,in order to improve the classification ability of neural networks,reduce parameter redundancy and improve training efficiency,a model of DenseNet with attention mechanism is proposed.The network can acquire important information of the image through the added attention mechanism,solve the data sensitivity problem,and improve the overall performance of the network.The accurate rates in Leafsnap tree leaf public data sets and SVHN of public data sets achieve 91.25%and 98.27%,respectively.The effect of DenseNet with attention mechanism is superior to other network model of neural network classification.

关 键 词：图像分类 卷积神经网络 密集神经网络 注意力机制 树种识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]