Abstract: When the shell is mixed with the tea seed, the traditional mechanical separation will still be doped with the shell, and the cleaning rate needs to be improved. In this paper, by comparing different layers of ResNet, it is found that in the current shell seed experiment samples, ResNet18 has the lowest average training time of each iteration compared with other models, and the highest average accuracy of verification set, which is more superior to other CNN classification models. In order to further improve the sorting efficiency, the attention mechanism was introduced into the ResNet18 network and the results showed that the average time per iteration in the training process of SE-ResNet18 model decreased from 1.31 s to 1.13 s by shortening 0.18 s, and the average accuracy of the validation set was 98.88% by improving 1.4 percentage points, compared with the model before improvement, and the overall accuracy of the test set was 98.43%, which was 13 percentage points higher than that of the original model, indicating that the use of ResNet18 model combined with the attention mechanism was feasible for the sorting of oil tea fruit shells and tea seeds, providing a new theoretical basis and thinking direction for the sorting method of camellia fruit shells and seeds.

Key words: camellia fruit, deep learning, sorting, SE-ResNet18 model, attention mechanism

摘要： 油茶果脱壳后果壳与茶籽混合在一起，采用传统的机械分选仍会出现掺杂果壳的情况，清选率有待提高。比较ResNet不同层数模型，发现在当前壳籽实验样本下ResNet18与其他模型相比每次迭代的平均训练时间最少，并且验证集平均准确率最高，同时均优于其他CNN分类模型。为进一步提升分选效率，在ResNet18网络中引入注意力机制，结果表明，SE-ResNet18模型与改进前的模型相比，训练过程中每次迭代的平均时间由1.31 s下降到1.13 s，缩短018 s，验证集平均准确率为98.88%，提升1.4个百分点。经过测试后得出，测试集整体准确率为98.43%，与原模型相比提升1.3个百分点，说明使用ResNet18模型结合注意力机制的方法在油茶果果壳与茶籽的分选上是可行的，为油茶果在分选方法提供一种新的理论基础与思考方向。

关键词: 油茶果, 深度学习, 分选, SE-ResNet18模型, 注意力机制