RTR-CNN: Rotated tray region selection and young seedling health status detection by CNN in greenhouse seedbed

Abstract

The use of mobile inspection transplanter machines in greenhouse seedling cultivation reduces reliance on manual labor, a critical step in vegetable production. Detecting plug seedlings and assessing their health are essential for automating seedling transplantation. In this study, a two-stage detection algorithm, RTR-CNN, was developed to locate and identify multiple rotating trays at various angles on greenhouse seedbeds and to evaluate the growth status of young seedlings. Focusing on a seedling tray with 200 cells, matching templates were designed based on the trays’ binary morphological features to detect trays at different angles. After locating trays, each cell was segmented and geometrically corrected before being fed into the backbone network for classification. Masked generative distillation was applied to optimize the network for edge computing hardware. Using the WideResNet101_2 teacher network, the deployed model achieved an accuracy of 88.80 %, with a binary classification accuracy of 98.95 %. Deployment on the lightweight MobileNetV2 network achieved an accuracy of 87.45 % through distillation learning, with a binary classification accuracy of 99.15 % at a processing speed of 159.30 fps with a batch size of 200. Distillation learning improved the training accuracy by 0.95 % and binary classification accuracy by 0.5 % compared to models trained without this method.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.