基于卷积神经网络的玉米种子质量检测

Maize Seed Quality Detection Based on Convolutional Neural Network

为提高玉米种子质量检测的精准率,以不同质量的单颗玉米种子为研究对象,提出一种基于卷积神经网络的玉米种子质量检测方法。通过旋转、镜像、明暗变换、添加噪声等方法对现有数据集进行扩充;利用调整对比度和Blob分析方法增强图像特征;选用卷积神经网络(CNN)的Resnet50分类模型与Squeezenet模型分别进行试验,将数据集按不同的比例分组,通过训练实现玉米种子按照好、坏、杂质的分类检测。最终对六组实验的综合评价指标对比得出:数据集按9:1的训练测试比例分配,选用Resnet50模型训练效果最好,平均综合指标达到96.45%。

In order to improve the accuracy of maize seed quality detection,a convolutional neural network based maize seed quality detection method was proposed with single maize seed of different quality as the research object.The existing data set was extended by rotation,mirror image,shading and noise addition.Contrast adjustment and Blob analysis were used to enhance image features.Resnet50 classification model of convolutional neural network(CNN)and Squeezenet model were selected to conduct experiments respectively.Data sets were grouped in different proportions,and the classification detection of maize seeds according to good,bad and impurity was realized through training.Finally,by comparing the comprehensive evaluation indexes of the six groups of experiments,it was concluded that when the training and testing sets were allocated in a 9:1 ratio,the Resnet50 model had the best training effect and the average comprehensive index(F1)was 96.45%.