Authors

Abstract

The germination of castor seeds was affected by different damage forms after shelling. Traditional methods could not express the change of mechanical damage characteristics on the surface of castor seeds. In the study, an improved migration learning algorithm for castor seed damage classification was adopted. The convolution kernel size of the first convolutional layer of the AlexNet model was modified, part of the convolutional layer was divided into two layers to increase the depth of the convolution model. Then a multi-scale convolution kernel was added to extract the damage characteristics of castor seeds. The results showed that combined with the hyperparameter optimization of convolutional layer stratification and the AlexNet model,the classification effect was improved. The average test accuracy was 98.10%. After the addition of multi-scale convolution, the average test accuracy was improved by 0.57%. The results show that the classification accuracy of cracked castor seeds is 71%, and the classification accuracy of castor seeds with missing shells is 63%. The classification accuracy of whole castor seeds is 67%. The verification of damage identification device for castor seeds was developed to verify the correctness of the algorithm. This study provided a theoretical and convolutional network model supported for the development of an online real-time damage classification detection system for castor seeds.

Abstract in English

蓖麻种子发芽率主要是脱壳后不同形式的损伤引起。传统方法无法表达蓖麻籽表面机械损伤特征的变化。 本研究采用了一种改进的蓖麻种子损伤分类迁移学习算法，修改了AlexNet模型第一个卷积层的卷积内核大小。卷积层被分成两层，以增加卷积模型的深度。添加一个多尺度卷积核，以提取蓖麻种子的损伤特征。 结果表明，结合卷积层分层的超参数优化和AlexNet模型，分类效果得到了改善。 添加多尺度卷积后，平均测试精度提高了0.57%。 实验结果显示，开裂蓖麻种子的分类准确率为71%，缺失壳蓖麻种子的分类准确率为63%。 整个蓖麻种子的分类准确率为67%。 本研究为开发蓖麻种子的在线实时损害分类检测系统提供了理论支持。