Transient multi-indicator detection for seedling sorting in high-speed transplanting based on a lightweight model

Abstract

Quality sorting of plug seedlings is an important part of factory nurseries, and deep learning is starting to become a core technology in this field. In response to the key problems of existing sorting systems with a single perspective and insufficient detection indicators, this paper innovatively proposes a lightweight multi-indicator detection model for plug seedlings, and an integrated transient sorting system for high-speed transplanting is designed. The new YOLOv5s model is composed of the ShuffleNet-V2 backbone, a channel attention mechanism neck, and a head block, and it can effectively extract image features of different seedling qualities. We built a dataset for pepper and tomato plug seedlings, and the experimental results showed that the model was highly accurate in detecting no seedlings, weak seedlings, damaged seedlings and strong seedlings, with an mAP of 94.23 %. Finally, in the verification of different operating speeds, the transplanting-sorting-replanting multifunctional system had the best integrated efficiency of 4200 plants per hour, and the average treatment time for single plug seedlings was only 0.86 s. Our method effectively can improve the performance of seedling quality sorting systems and provide technical support for factory nurseries.