自组织CNN建模识别耐热钢金相组织研究

Research on Self⁃organized CNN Modeling to Identify Metallographic Structure of Heat⁃resistant Steel

耐热钢金相组织结构复杂,传统的图像分析方法特征提取困难,难以对其进行准确的自动识别,而人工识别易受主观因素影响,导致识别精度波动大,结果重复性差。卷积神经网络(Convolutional neural networks,CNN)能够从原始图像中提取复杂的特征,但是CNN建模需要的训练与拓扑超参数选择和优化困难。本工作利用基于超参数组合计算资源分配的Hyperband算法来优化CNN模型的超参数,克服了网格搜索、随机搜索及贝叶斯优化效率低、计算资源消耗量大以及优化不稳定等问题,实现自组织CNN建模。基于Hyperband算法优化得到33层CNN模型,进行训练与仿真,并结合混淆矩阵对模型的识别结果进行评价。结果表明,所建模型对耐热钢金相组织识别的准确率、精确度、灵敏度、特异度的均值分别为94.2%、94.1%、94.2%和98.1%,表明模型具有较高的泛化能力,能够较为准确地识别金相组织,为复杂金相组织的智能识别提供新方法。

Heat-resistant steel has complex metallographic structure, hence it is difficult to extract features using traditional image processing methods to accurately achieve auto recognition. Manual recognition is susceptible to subjective factors and thus resulting in poor accuracy and repeatability. Convolutional neural networks(CNN) can extract complex features from raw images, only with the problem of difficulty in model training and selection and optimization of topological hyperparameters. In this work, a Hyperband algorithm, which calculates resource allocation based on hyperparameter sets, was used to optimize hyperparameter in CNN model and finally the self-organizing modelling was achieved. Relevant problems were solved such as low efficiency and large consumption in computing resources due to grid searching, random searching and Bayesian optimization as well as optimization instability. Based on above-mentioned Hyperband algorithm optimization, a 33 layers’ CNN model was constructed, trained, simulated and its recognition performance evaluated with confusion matrix analysis. Results showed that, in recognition of metallographic structure for heat-resistant steel, an average accuracy of 94.2%, an average precision of 94.1%, an average sensitivity of 94.2 and an average specificity of 98.1% were obtained after using the established model. Our model demonstrated a good generalization ability and could be used to recognize metallagraphic structure accurately. This approach provided a new method in intelligent recognition of complex metallagraphic structures.