Abstract Human labor still play an important role in cane sugar crystallization process. Automation control is essential to reduce human labor. An accurate image classification system is the basis for automation control of the cane sugar crystallization process. This paper builds a deep learning framework based on deep convolutional neural networks (DCNNs) to classify cane sugar crystallization image of cane sugar crystallization process for sugar factory. Different networks were trained on a large image data set obtained from a sugar batch crystallizer. Based on the data set, the established model was used to classify cane sugar crystallization image. The classification accuracy of the proposed model reached 0.901. The confusion matrix of the InceptionResNetV2 model indicates classification accuracy of between 0.83 and 0.99 are achieved in classifying cane sugar crystal images from a cane sugar factory into 5 categories. This provides a promising means for the future development of monitoring systems using image. The proposed DCNNs model was compared against other models, such as, Inception-V3, ResNet50, and a simple DCNNs. The experimental results showed that the deep learning framework outweighs other models and can serve as a benchmark of monitoring cane sugar crystallization using DCNNs in sugar industry.

中文翻译：

---

使用深度神经网络监测糖结晶

摘要 人工劳动在蔗糖结晶过程中仍起着重要作用。自动化控制对于减少人工劳动至关重要。准确的图像分类系统是蔗糖结晶过程自动化控制的基础。本文构建了一个基于深度卷积神经网络（DCNNs）的深度学习框架，对糖厂蔗糖结晶过程的蔗糖结晶图像进行分类。在从糖分批结晶器获得的大型图像数据集上训练不同的网络。基于数据集，建立的模型用于对蔗糖结晶图像进行分类。所提出模型的分类准确率达到了0.901。InceptionResNetV2 模型的混淆矩阵表明分类精度在 0.83 和 0 之间。将蔗糖厂的蔗糖晶体图像分为5类，达到99个。这为使用图像的监控系统的未来发展提供了有希望的手段。将提出的 DCNN 模型与其他模型进行了比较，例如 Inception-V3、ResNet50 和简单的 DCNN。实验结果表明，深度学习框架优于其他模型，可以作为糖业中使用 DCNN 监测蔗糖结晶的基准。