| 基于Res2Net和双线性注意力的番茄病害时期识别方法 |
| |
| 引用本文： | 贾兆红,张袁源,王海涛,梁栋.基于Res2Net和双线性注意力的番茄病害时期识别方法[J].农业机械学报,2022,53(7):259-266. |
| |
| 作者姓名： | 贾兆红 张袁源 王海涛 梁栋 |
| |
| 作者单位： | 安徽大学互联网学院 |
| |
| 基金项目： | 国家自然科学基金项目（71971002）、安徽省重大专项（202003a06020016）和安徽省教育厅科学研究项目（YJS20210029） |
| |
| 摘 要： | 针对番茄叶片型病害在早晚期具有类内差异大、类间差异小的特点，常规神经网络对此类病害的分类效果不佳的问题，提出了基于Res2Net和双线性注意力的番茄病害时期识别方法，通过多尺度特征和注意力机制，提高网络的细粒度表征能力。首先，提出EFCA通道注意力模块，在不降维的基础上，使用二维离散余弦变换代替全局平均池化，以减少常规通道注意力获取时的信息丢失。其次，在外积之后加入最大池化和concat操作，避免双线性融合后因维度过高导致的特征冗余。在7种不同种类和14种不同程度病害番茄叶面型病害数据集实验中，本文方法分类准确度分别为98.66%和86.89%。
|
| 关 键 词： | 神经网络 细粒度分类 番茄病害时期 注意力机制 双线性卷积神经网络 |
| 收稿时间： | 2021/7/16 0:00:00 |
Identification Method of Tomato Disease Period Based on Res2Net and Bilinear Attention Mechanism |
| |
| JIA Zhaohong,ZHANG Yuanyuan,WANG Haitao,LIANG Dong.Identification Method of Tomato Disease Period Based on Res2Net and Bilinear Attention Mechanism[J].Transactions of the Chinese Society of Agricultural Machinery,2022,53(7):259-266. |
| |
| Authors: | JIA Zhaohong ZHANG Yuanyuan WANG Haitao LIANG Dong |
| |
| Institution: | Anhui University |
| |
| Abstract: | Tomato leaf-type diseases have the characteristics of large intra-class differences and small inter-class differences in the early and late stages. The conventional neural network is not effective in classifying such diseases. Therefore, based on the fine-grained weakly supervised classification method, a Res2Net bilinear attention network, combining the bilinear model and attention mechanism, was proposed. The fine-grained representation ability was improved through extracting multi-scale features and combining the attention mechanism. First of all, for the problem of information loss in the process of conventional channel attention acquisition, the EFCA channel attention module was proposed. On the basis of no dimensionality reduction, two-dimensional discrete cosine transform was used instead of global average pooling to avoid some features from being lost in downsampling. Secondly, by adding the maximum pooling after the outer product, and the concat module designed by drawing on the shortcut idea in the residual network, the problem of redundant features caused by the excessively high dimensionality after bilinear fusion was solved. The obtained classification accuracies of the proposed model on the data set with 7 types and 14 different degrees of tomato leaf type diseases were 98.66% and 86.89%, respectively. |
| |
| Keywords: | neural network fine grained classification tomato disease period attention mechanism bilinear convolutional neural network |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
