基于卷积神经网络与迁移学习的碳钢石墨化自动评级研究

Automatic Evaluation Study of Carbon Steel Graphitization Based on Convolutional Neural Network and Transfer Learning

摘要:

为实现碳钢石墨化的智能化评级,基于卷积神经网络与迁移学习的方法构建了碳钢金相图像的自动分类模型。首先通过几何变换和像素调整的数据增强方法建立了碳钢石墨化图像数据集。然后采用统一扩展网络宽度、深度和分辨率方式来协调精度与效率的轻量级EfficientNet网络作为主干特征提取网络,构建碳钢石墨化图像评级模型,并在训练阶段利用迁移学习与参数微调的方法来提高模型的训练效率。最后使用测试数据集对模型的分类精度与复杂度进行了验证实验,结果表明该模型能快速准确的对碳钢石墨化程度进行自动评级,在仅需12MB内存的情况下,便可实现97.01%的评级准确率,单幅金相图像的平均检测时间也仅需10.27ms,满足现场检测的精度与实时性要求。

关键词:碳钢石墨化;金相图像;卷积神经网络;迁移学习;自动评级

Abstract:

In order to realize the intelligent evaluation of carbon steel graphitization, an automatic classification model of carbon steel metallographic images is constructed based on convolutional neural network and transfer learning. The carbon steel graphitization image dataset was firstly established by the data enhancement methods of geometric transformation and pixel adjustment. Then the lightweight EfficientNet network that uniformly expands the network width, depth and resolution to coordinate accuracy and efficiency was used as the backbone feature extraction network to construct a carbon steel graphitization image evaluation model, and transfer learning and parameter fine-tuning methods were used in the training phase to improve the training efficiency of the model. Finally, a test data set was used to verify the classification accuracy and complexity of the model. The results show that the model can quickly and accurately grade the degree of graphitization of carbon steel automatically. With only 12 MB of memory, it can achieve a 97.01% accuracy, and an average detection time of only 10.27 ms for a single metallographic image, which meets the accuracy and real-time requirements of on-site detection.

Key words:carbon steel graphitization; metallographic image; convolutional neural network; transfer learning; automatic evaluation