| 本文已被：浏览 1784次 下载 925次 |
码上扫一扫！ |
| 基于卷积块注意力胶囊网络的小样本水稻害虫识别 |
|
曾伟辉1,2,3,唐欣1,胡根生1*,梁栋1
|
|
|
| (1.安徽大学 农业生态大数据分析与应用技术国家地方联合工程研究中心, 合肥 230601;2.安徽大学 互联网学院, 合肥 230039;3.科大国创软件股份有限公司, 合肥 230088) |
|
| 摘要: |
| 针对真实复杂背景下小样本水稻害虫识别模型泛化能力弱，易受复杂背景干扰以及重要特征表达能力不强等问题，提出了一种基于卷积块注意力胶囊网络的小样本水稻害虫识别方法。采用数据增强的方法扩充数据，以提高模型的泛化能力同时预防过拟合；利用GrabCut算法去除图像中的复杂背景，减小复杂背景对水稻害虫识别的干扰；将空间注意机制和通道注意机制相结合的卷积块注意模块(Convolutional block attention module，CBAM)引入到胶囊网络中，提高模型对水稻害虫特征的表达能力，使模型关注重要特征，抑制不必要的特征。其中胶囊网络主要用来更加敏锐地发现小样本图像中水稻害虫的相对位置和角度等信息。结果表明:在对复杂背景下小样本水稻害虫的识别时，本研究方法准确识别率达99.19%，优于支持向量机(Support vector machine，SVM)、k近邻(k-nearest neighbors，kNN)等浅层网络方法，也优于VGG16、GoogLeNet以及Mobilenet等深度网络方法，可实现复杂背景下的小样本水稻害虫的准确识别。 |
| 关键词: 水稻害虫 害虫识别 卷积神经网络 胶囊网络 注意力机制 |
| DOI：10.11841/j.issn.1007-4333.2022.03.08 |
| 投稿时间：2021-07-21 |
| 基金项目:安徽省自然科学基金面上项目(2108085MC95)；安徽省科技重大专项(202003a06020016)；安徽省高校自然科学研究项目(KJ2020ZD03, KJ2020A0039)；农业生态大数据分析与应用技术国家地方联合工程研究中心开放课题(AE202004) |
|
| Rice pests recognition with small number of samples based on CBAM and capsule network |
|
ZENG Weihui1,2,3,TANG Xin1,HU Gensheng1*,LIANG Dong1
|
| (1.National Engineering Research Center for Agro-Ecological Big Data Analysis & Application, Anhui University, Hefei 230601, China;2.School of Internet, Anhui University, Hefei 230039, China;3.Guochuang Software Co., Ltd., University of Science and Technology of China, Hefei 230088, China) |
| Abstract: |
| Aiming at the problems of weak generalization ability, vulnerable to complex background interference and weak expression ability of important features of rice pest recognition model with small number of samples under real complex background, a rice pest recognition method with small number of samples based on convolution block attention capsule network is proposed. The data are expanded by data augmentation method to improve the generalization ability of the model and prevent over-fitting. GrabCut algorithm is used to remove the complex background in the image and reduce the interference of complex background on the identification of rice pests, convolutional block attention module combined with spatial attention mechanism and channel attention mechanism is introduced into the capsule network to improve the expression ability of the model to the features of rice pests, so that the model is designed to focuse on important features and suppress unnecessary features. The capsule network is mainly used to find the relative position and angle of rice pests with small sample size problems more sensitively. The results show that for the identification of rice pests with small sample size problems, the accuracy of the proposed method can reach 99. 19%, which is better than that of support vector machine, k-nearest neighbors, VGGNet, GoogLeNet and Mobilenet. The proposed method can achieve accurate identification of rice pests with small number of samples under the complex rice background. |
| Key words: rice pest pest identification convolutional neural networks capsule network attention mechanism |
