Morphological recognition of rice seedlings based on GoogLeNet and UAV image

Objective  In view of the current situation that the quality of transplanting is mainly based on manual observation and random sampling in China, it is proposed to use the convolutional neural network GoogLeNet to recognize the morphology of rice seedlings.

Method  Firstly, clear and intact images of rice seedlings were obtained by UAV aerial photography at low altitude. Data sets of floating seedlings, damaged seedlings and qualified seedlings were made by cutting and marking. Then, based on the GoogLeNet structure training data, the optimal network recognition model was obtained. Finally, the image classification experiment of seedlings per hole was carried out, and compared with traditional image classification algorithms (SVM, BP neural network).

Result  Under the condition of using the same samples, the seedling morphology recognition method based on GoogLeNet completed the judgment and classification was faster and more accurately. The average accuracy of seedling morphology recognition was 91.17%, and the average detection time was 0.27 s. Compared with SVM and BP neural network, the average classification accuracy increased by 21 and 13 percentage points respectively, and the detection time was shortened by 1.09 and 0.58 s respectively.

Conclusion  This study can provide the relevant support for evaluation of rice transplanting quality.