基于卷积神经网络的牧草种子图像识别  

Convolutional neural network-based image recognition of forage seeds

作　　者：何欣[1] 王雪萌 张涵 宋瑞 田沛鑫 刘萍[1] 毛培胜[1] 贾善刚 HE Xin;WANG Xuemeng;ZHANG Han;SONG Rui;TIAN Peixin;LIU Ping;MAO Peisheng;JIA Shan’gang(College of Grassland Science and Technology,China Agricultural University,Beijing 100191,China)

机构地区：[1]中国农业大学草业科学与技术学院,北京100091

出　　处：《草业科学》2022年第11期2338-2349,共12页Pratacultural Science

基　　金：现代农业产业技术体系(CARS-34);中国农业大学基本业务费(2019TC257、2020TC189)。

摘　　要：传统的灰度图像处理方式会降低数据集的复杂性,进而降低模型识别相似种子的准确率。因此,将卷积神经网络中传统的图像灰质化处理方法改进为归一化典型判别分析(nCDA)算法与卷积神经网络(CNN)相结合的方法(CNN-nCDA),该方法基于多光谱仪采集的种子图像信息,以深度学习框架TensorFlow为基础,利用卷积神经网络构建种子识别算法,可区分高相似度牧草种子。结果表明,仅依靠传统灰度图难以区分形态相似的种子(62.11%~72.5%),而采用CNN-nCDA策略对不同类别种子分类的准确率可达100.0%,优于单独使用nCDA(90.0%~100.0%)、线性判别分析(LDA)(97.3%~100.0%)、支持向量机(SVM)(92.4%~97.5%)的准确率。综上所述,多光谱成像技术中的nCDA算法与卷积神经网络相结合技术,具有较高的校准和验证能力,对现场快速筛选种子具有良好的应用前景。The traditional method of seed identification based on grayscale image processing reduces image data complexity and prediction accuracy.This study improved this method to a more efficient method named convolutional neural networks combined with normalized canonical discriminant analysis(CNN-nCDA),which combines two techniques,nCDA and CNN.CNN-nCDA employs the deep learning framework TensorFlow to distinguish forage seeds based on multispectral imaging,and its seed recognition algorithm is conducted by CNN.The results showed that for seed identification with similar morphology,the accuracy rate of the CNN-nCDA strategy could reach 100.0%.This is a vast improvement over traditional grayscale image processing(62.11%~72.5%),nCDA(90.0%~100.0%),linear discrimination analysis(97.3%~100.0%),and support vector machine(92.4%~97.5%).In conclusion,the CNN-nCDA strategy has high calibration and verification capabilities and holds promise for the purpose of rapid seed identification.

关 键 词：TensorFlow 卷积神经网络 多光谱 种子识别 图像识别 牧草种子 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] S54[农业科学—作物学]