基于ECA-FV-CNN的水稻单籽粒质量分级方法

Method for Single Rice Grain Weight Grading Based on ECA-FV-CNN

为解决传统水稻质量分级依靠人工分拣,工作量大、错误率高、分级标准不严格等问题,本文提出一种基于ECA改进的双流卷积神经网络模型对水稻单粒质量分级进行研究。首先,获取每组水稻单籽粒(本文以7颗水稻单籽粒为1组)正视和俯视图像,对于5种简单的监督模型(朴素贝叶斯、决策树、随机森林、最邻近结点算法、支持向量机)、基于遗传算法和投票机制优化的模型(GA-SVM)、集成模型(RF+GA-SVM),通过图像预处理轮廓检测分离出单籽粒图像,利用颜色矩、LBP(Local binary pattern)和Canny算子提取籽粒颜色、纹理和边缘特征,并采用PCA(Principal component analysis)降维后进行训练;而对于单流卷积神经网络模型、双流卷积神经网络模型(FV-CNN)以及本文提出并构建的基于ECA改进的双流卷积神经网络模型(EA-FV-CNN),则使用预处理后的图像进行训练。将上述多种模型进行对比分析,发现基于ECA改进的双流卷积神经网络模型性能最好,其在单粒质量三分级、四分级和五分级准确率分别达94.0%、92.3%和71.0%。实验结果表明,使用基于ECA改进的双流卷积神经网络模型能够提高水稻单粒质量的分级精度,弥补传统方法的不足,规范籽粒筛选分级标准。

Aiming to solve the problems that traditional grain weight classification depends on manual sorting,such as heavy workload,high error rate and lax classification standard,an improved two-stream convolutional neural network model was proposed based on ECA to classify rice by single grain weight.Firstly,images of each group of rice(a group consists seven single rice grains)were taken from two different perspectives:front view and top view.For five traditional supervised models(naive Bayes,decision tree,random forest,K-nearest neighbor,support vector machine),voting mechanism optimization based on genetic algorithm(GA)(GA-SVM)and integrated model(RF+GA-SVM),single grain images were separated through image preprocessing and contour detection.Color moment,local binary pattern(LBP)and Canny operator were used to extract grain color,texture and edge features.And then through principal component analysis(PCA),the principal features were extracted to train each model.For the constructed single-stream convolutional neural network model,two-stream convolutional neural network model(FV-CNN)and the improved two-stream convolutional neural network model were proposed based on ECA(ECA-FV-CNN),the pre-processed images were divided into training set,verification set and test set according to the ratio of 6∶2∶2,and data enhancement were carried out for each data set,and then the models were trained.By comparing and analyzing the above models,the traditional machine learning model,RF+GA-SVM,had the best effect,but its highest accuracy was only 72%when the single grain weight was set for three-graded.Experimental verification showed that the ECA-FV-CNN model proposed had the best performance,and its accuracy for the single grain weight classification of three-graded,four-graded and five-graded reached 94.0%,92.3%and 71.0%,respectively.However,the accuracies of single-stream convolutional neural network model and FV-CNN model for single grain weight grading were 92.7%,91.1%,61.1%and 93.0%,91.6%,65.6%,respectively.The grading effect of FV-CNN model was better than that of singlestream convolutional neural network model in three experiments,which showed that the two-branch network training was better than that of single-branch rice single grain weight grading.The accuracy of ECA-FV-CNN model in three grading experiments was 16.2%higher than that of single-stream convolutional neural network model and 8.2%higher than that of FV-CNN model.The results showed that the introduction of ECA module was effective for rice single grain classification,and the improved two-stream convolutional neural network model based on ECA can improve the classification accuracy of rice single grain weight,and the classification of rice single grain weight can be achieved by using computer vision technology,making up for the shortcomings of traditional methods,and improving the classification standard of grain screening.