High-throughput maize seed ears sorting through structural re-parameterisation classification model and multi-channel sorting system

Abstract

Sorting maize ears is critical to ensure the quality and purity of maize seeds. However, sorting is the most time-consuming and labour-intensive process, requiring an intelligent approach and automated device for high throughput. As a first step, this paper designs an automated image acquisition system based on sensor-triggered cameras to decrease the time and cost of producing ear dataset. Secondly, a structural re-parameterisation classification model is proposed, and this model achieved "deep training, shallow reasoning" by complicating the structure only during training and re-parameterising it to the original structure for inference, thus decoupling the complexity and the inference time. Then, the model trained on the PC is deployed in Nvidia TX2 NX, and the TensorRT is used to nearly double the model inference speed. This model reduces the demand for computational resources, thereby saving development costs. Subsequently, a multi-channel sorting system is developed based on the classification model, machine vision, sensors, and automatic control technologies to realise high-throughput ear sorting. Finally, the sorting experiments show that the average model’s classification accuracy on the dataset of multiple varieties is 97.25 %, demonstrating a good generalisation ability. The average sorting accuracy of the sorting system is 96.25 %, and the processing capacity of a single working channel is about 300 kg h⁻¹, preparing the foundation for the subsequent development of an intelligent maize seed ear sorting device. This study can provide a reference for the sorting of other elongated crops.