摘要
传统的花卉图像分类都是基于人工手动选择单一特征或者多特征融合再分类,这种方法普遍存在精度低、成本高、泛化能力弱等缺陷。针对目前深度学习在细粒度图像分类中的应用,提出一种基于残差网络、实现端到端的花卉图像分类方法。首先以ResNet18为基础模型;其次将全卷积结构的思想应用于网络模型中,将ResNet18的全连接层替换成卷积层以优化网络模型;最后在优化后的ResNet18中融入混合域注意力机制,由Softmax层进行分类。本文选取了Oxford17flowers和Oxford102flowers两个花卉图像数据集做对比试验,与前人的花卉图像分类方法对比,本文的方法取得了理想的效果,在Oxford17和Oxford102上分别取得了99.26%以及99.02%的正确率。提出了一种基于注意力的残差结构改进方法,相较于前人的花卉图像分类方法,该方法能够更有效地提取关键信息的特征,抑制干扰区域的信息,对花卉图像分类具有显著性效果,适用于细粒度图像分类。
Traditional flower classification usually obtains a significant area by image segmentation and manually extracting or choosing suitable features. This method generally lacks in high precision,low cost and strong generalization ability. Aiming at the application of deep learning in fine-grained image classification,an end-to-end method of flower classification based on residual network is proposed.Firstly,the method uses ResNet18 as the basis model. Secondly,this paper applies the idea of full convolution structure to the network model.The convolution layer replaces the fully connected layer of ResNet18 to optimize the network model(Res18_conv). Finally,the residual attention network is incorporated into the optimized ResNet18(Res18_conv_a). The residual attention network studies a mixed domain,focusing on the spatial domain as well as the channel domain. After the entire network extracts all flower features,the flower is classified by the softmax function. This paper selected two data sets of Oxford17 flowers and Oxford102 flowers to verify the proposed image classification method based on improved residual network. The method of this paper has achieved the desired effect compared with the previous flower image classification method. For Oxford17 data set,the accuracy of Res18_conv_a is 99.26%. For Oxford102 data set,the accuracy of Res18_conv_a is 99.02%. This paper proposes an improved residual structure based on attention mechanism. Compared with the predecessor’s floral image classification method,this method can extract the characteristics of key information more effectively and suppress the information of the interference area. This method has a significant effect on the classification of flower images,which is suitable for fine-grained image classification.
作者
裴晓芳
张杨
PEI Xiaofang;ZHANG Yang(Binjiang College,Nanjing University of Information Science&Technology,Nanjing 210044,China;Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology,Nanjing University of Information Science&Technology,Nanjing 210044,China;College of Electronic and Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China)
出处
《电子器件》
CAS
北大核心
2020年第3期698-704,共7页
Chinese Journal of Electron Devices
基金
南京信息工程大学滨江学院课题项目(2019bjyng006)。
关键词
图像分类
花卉识别
残差网络
全卷积结构
注意力机制
image classification
flower recognition
residual network
full convolution structure
attention mechanism
