Abstract: In order to realize the low cost，efficient and convenient variety identification of coated maize seed，a dataset was constructed based on 23 100 double‑sided images of 18 varieties and 4 colors of coated maize seeds collected by smartphone，and the lightweight convolutional neural network models ShuffleNetV2，MobileNetV3，MobileViT，MobileOne，RepGhostNet and the integrated models based on the above models were used to identify coated maize seed variety.The results showed that the identification accuracy and the comprehensive performance of the five single models were high.The identification accuracies were 98.48%，98.23%，98.44%，98.23% and 98.01%，respectively.The model sizes were 1.55，4.96，4.42，6.97 and 4.19 MB，respectively. The inference speeds were 106，94，84，212 and 94 f/s，respectively.The identification accuracy of the integrated models was higher than that of the single models，and the identification accuracy of the integrated model composed of ShuffleNetV2 and MobileViT was 99.22%.The analysis found that the false identification only occurred in the varieties of the same color coated seeds，and as the number of varieties of the same color coated seeds increased，the model’s identification accuracy had a downward trend.

Key words: Maize, Coated seed, Variety identification, Visible light, Deep learning

摘要： 为实现玉米包衣种子品种低成本、高效便捷识别，基于智能手机采集的18个品种4种颜色的23 100张玉米包衣种子双面图像构建数据集，采用轻量级卷积神经网络模型ShuffleNetV2、MobileNetV3、MobileViT、MobileOne、RepGhostNet和基于上述模型的集成模型分别进行玉米包衣种子品种识别。结果表明，5种单一模型均具有较高的识别准确率和综合性能，识别准确率分别为98.48%、98.23%、98.44%、98.23% 和98.01%，模型大小分别为1.55、4.96、4.42、6.97、4.19 MB，推理速度分别为106、94、84、212、94 f/s。集成模型相比单一模型具有更高的识别准确率，其中，ShuffleNetV2和MobileViT组成的集成模型识别准确率达到99.22%。分析发现，品种误识别仅发生在相同颜色包衣种子品种之间，并且随着相同颜色包衣种子品种数量增多，模型对该颜色包衣种子的识别准确率有下降的趋势。

关键词: 玉米, 包衣种子, 品种识别, 可见光, 深度学习