Abstract
Abstract. With the development of precision agriculture, the selection of maize kernels has gained more importance in scientific research and practical significance in agricultural production. In this study, the deep learning technology of machine vision was used to select maize kernels, solving the problems of previous maize kernel selection for specific sorting problems, the cumbersome process of artificial feature modeling, the problem of a small number of features, and the challenge of limited data. First, the maximum size of a model based on convolutional neural networks (CNNs) that could run under finite hardware conditions was determined by experiments. Four different network models (Faster R-CNN, Model 1.0, Model 2.0, and Model 3.0) were then designed and trained using a data set of maize kernels. Finally, the accuracy of the models was verified by comparison test, and the detection results of the models were analyzed according to their precision, recall, FPR, F1, precision-recall curve, average precision (AP), mean average precision (mAP), and detection speed. The results show that for the validation set not used for training, Model 1.0 had the highest average recall rate of 98.42% among the four models. Without taking into account the identification of the removed kernels, only excellent maize kernels were identified, and the mAP of Model 1.0 was as high as 97.27%. Moreover, Model 1.0 requires less computer resources, and its computer hardware requirement is lower. The precision, recall, and F1 value of Model 2.0 were increased by 3.73%, 3.55%, and 3.79%, respectively, and the false positive rate of Model 2.0 was reduced by 1.31% on average compared with the Faster R-CNN model. By comparing Model 1.0, Model 2.0, and Model 3.0, it was found that the overall performance of Model 2.0 was best. The size of the network model has an effect on the accurate selection of maize kernels, and a moderate-size model is the best. This study laid a good foundation for the further application of deep learning technology in the real-time sorting of maize kernels and additional applications in the field of agriculture. Keywords: Convolutional neural networks, Deep learning, Maize kernel, Selection, Visualization.
Abstract from the paper’s OpenAlex record, shown for indexing and research. Copyright remains with the authors or the publisher. Read at the source ↗
The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.
Claims 9
Claim · model-written summary
Model 1.0 had the highest recall, 98.42%.
Not reportedTaxon
Verbatim
e0“Model 1.0 had the highest average recall rate of 98.42% among the four models”
Claim · model-written summary
Model 1.0 reached 97.27% mAP for excellent kernels only.
Zea maysOrganisms pictured · Representative pictures of the matched species, not figures from the paper.
Verbatim
e0“only excellent maize kernels were identified, and the mAP of Model 1.0 was as high as 97.27%”
Claim · model-written summary
Model 1.0 required fewer computer resources.
Verbatim
e0“Model 1.0 requires less computer resources, and its computer hardware requirement is lower”
Claim · model-written summary
Model 2.0 precision exceeded Faster R-CNN by 3.73%.
Verbatim
e0“The precision, recall, and F1 value of Model 2.0 were increased by 3.73%, 3.55%, and 3.79%, respectively, and the false positive rate of Model 2.0 was reduced by 1.31% on average compared with the Faster R-CNN model”
Claim · model-written summary
Model 2.0 recall exceeded Faster R-CNN by 3.55%.
Verbatim
e0“The precision, recall, and F1 value of Model 2.0 were increased by 3.73%, 3.55%, and 3.79%, respectively, and the false positive rate of Model 2.0 was reduced by 1.31% on average compared with the Faster R-CNN model”
Claim · model-written summary
Model 2.0 F1 exceeded Faster R-CNN by 3.79%.
Verbatim
e0“The precision, recall, and F1 value of Model 2.0 were increased by 3.73%, 3.55%, and 3.79%, respectively, and the false positive rate of Model 2.0 was reduced by 1.31% on average compared with the Faster R-CNN model”
Claim · model-written summary
Model 2.0 false positive rate was reduced by 1.31% versus Faster R-CNN.
Verbatim
e0“the false positive rate of Model 2.0 was reduced by 1.31% on average compared with the Faster R-CNN model”
Claim · model-written summary
Model 2.0 had the best overall performance among three variants.
Verbatim
e0“By comparing Model 1.0, Model 2.0, and Model 3.0, it was found that the overall performance of Model 2.0 was best”
Claim · model-written summary
Moderate model size gave the best kernel selection.
Organisms pictured · Representative pictures of the matched species, not figures from the paper.
Verbatim
e0“The size of the network model has an effect on the accurate selection of maize kernels, and a moderate-size model is the best”