Information Acquisition Method of Tomato Plug Seedlings Based on Cycle-Consistent Adversarial Network

Abstract

In order to solve the interference caused by the overlapping and extrusion of adjacent plug seedlings, accurately obtain the information of tomato plug seedlings, and improve the transplanting effect of automatic tomato transplanters, this study proposes a seedling information acquisition method based on Cycle-Consistent Adversarial Network (CycleGAN). CycleGAN is a generative unsupervised deep learning method, which can realize the free conversion of the source-domain plug seedling image and the target-domain plug label image. It collects more than 500 images of tomato plug seedlings in different growth stages as a collection image set; follows certain principles to label the plug seedling images to obtain a label image set, and uses two image sets to train the CycleGAN network model. Finally, the trained model is used to process the images of tomato plug seedlings to obtain their label images. According to the labeling principle, the correct rate of model recognition is between 91% and 97%. The recognition results show that the CycleGAN model can recognize and judge whether the seedlings affected by the adjacent seedling holes are suitable for transplanting, so the application of this method can greatly improve the intelligence level of the automatic tomato transplanters.