《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (8): 2593-2601.DOI: 10.11772/j.issn.1001-9081.2022060893
收稿日期:
2022-06-20
修回日期:
2022-09-05
接受日期:
2022-09-09
发布日期:
2022-09-22
出版日期:
2023-08-10
通讯作者:
陈金水
作者简介:
黄学雨(1970—),男,江西赣州人,教授,博士,主要研究方向:企业信息化、智能工厂基金资助:
Xueyu HUANG1,2, Huaiyu HE1, Huimin LIN1, Jinshui CHEN3,4()
Received:
2022-06-20
Revised:
2022-09-05
Accepted:
2022-09-09
Online:
2022-09-22
Published:
2023-08-10
Contact:
Jinshui CHEN
About author:
HUANG Xueyu, born in 1970, Ph. D., professor. His research interests include enterprise informatization, smart factory.Supported by:
摘要:
针对铜合金成分检测过程中产生的时滞问题,提出一种基于特征聚合的铜合金金相图分类识别方法。首先,在特征提取阶段,构建灰度共生矩阵(GLCM)和基于卷积注意力模块的残差网络(ResNet)模型分别提取图像的全局与局部特征;其次,在特征聚合阶段,将提取到的特征规范化后进行简单的级联;最后,在分类识别阶段,使用支持向量机(SVM)精确分类。实验结果表明,所提方法的准确率达到了98.963%、宏F1达到了98.996%,优于基于单特征的机器学习方法。可见,不同的方法提取的特征经过聚合后可以更全面地描述铜合金金相图的纹理及边缘信息,所提方法可以通过金相图识别不同铜合金,提升了识别的准确率,且具有良好的鲁棒性。
中图分类号:
黄学雨, 贺怀宇, 林慧敏, 陈金水. 基于特征聚合的铜合金金相图分类识别方法[J]. 计算机应用, 2023, 43(8): 2593-2601.
Xueyu HUANG, Huaiyu HE, Huimin LIN, Jinshui CHEN. Classification and recognition method of copper alloy metallograph based on feature aggregation[J]. Journal of Computer Applications, 2023, 43(8): 2593-2601.
数据集类型 | 不同成分的铜合金 | |||
---|---|---|---|---|
Cu-Cr-Zr | Cu-Fe | Cu-Fe-Cr | Cu-Fe-Mg | |
训练集 | 1 020 | 878 | 824 | 324 |
验证集 | 348 | 296 | 276 | 108 |
测试集 | 336 | 272 | 266 | 90 |
表1 预处理后的数据集分布情况
Tab. 1 Distribution of pre-processed dataset
数据集类型 | 不同成分的铜合金 | |||
---|---|---|---|---|
Cu-Cr-Zr | Cu-Fe | Cu-Fe-Cr | Cu-Fe-Mg | |
训练集 | 1 020 | 878 | 824 | 324 |
验证集 | 348 | 296 | 276 | 108 |
测试集 | 336 | 272 | 266 | 90 |
阶段 | 输出大小 | ResNet-CBAM |
---|---|---|
Stage0:Conv 1 | 224 | 3 |
Stage1:CBMA | 224 | 1 1 |
224 | 7 | |
Stage2:Conv2 | 224 | 3 |
Stage3:Conv3 | 112 | |
Stage4:Conv4 | 56 | |
Stage5:Conv5 | 28 | |
Stage6:CBMA | 28 | 1 1 |
28 | 7 | |
Stage7 | 1 | Average pool,1 024 -d fc |
表2 ResNet-CBAM网络参数
Tab. 2 ResNet-CBAM network parameters
阶段 | 输出大小 | ResNet-CBAM |
---|---|---|
Stage0:Conv 1 | 224 | 3 |
Stage1:CBMA | 224 | 1 1 |
224 | 7 | |
Stage2:Conv2 | 224 | 3 |
Stage3:Conv3 | 112 | |
Stage4:Conv4 | 56 | |
Stage5:Conv5 | 28 | |
Stage6:CBMA | 28 | 1 1 |
28 | 7 | |
Stage7 | 1 | Average pool,1 024 -d fc |
模型 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
ResNet50 | 95.746 | 96.151 | 95.324 | 95.735 |
ResNet101 | 96.784 | 97.025 | 97.375 | 97.199 |
DenseNet121 | 97.925 | 98.213 | 98.325 | 98.162 |
Xception | 96.887 | 97.174 | 96.875 | 97.024 |
ResNet-CBAM | 97.510 | 97.525 | 97.662 | 97.662 |
表3 测试集上不同卷积神经网络模型的性能比较 (%)
Tab. 3 Performance comparison of different convolutional neural network models on test set
模型 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
ResNet50 | 95.746 | 96.151 | 95.324 | 95.735 |
ResNet101 | 96.784 | 97.025 | 97.375 | 97.199 |
DenseNet121 | 97.925 | 98.213 | 98.325 | 98.162 |
Xception | 96.887 | 97.174 | 96.875 | 97.024 |
ResNet-CBAM | 97.510 | 97.525 | 97.662 | 97.662 |
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
GLCM-KNN | 81.327 | 81.988 | 79.831 | 79.995 |
GLCM-DT | 73.236 | 73.804 | 72.788 | 73.144 |
GLCM-SVM | 92.738 | 93.714 | 93.234 | 93.412 |
表4 测试集上基于GLCM的不同分类方法性能比较 (%)
Tab. 4 Performance comparison of different GLCM-based classification methods on test set
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
GLCM-KNN | 81.327 | 81.988 | 79.831 | 79.995 |
GLCM-DT | 73.236 | 73.804 | 72.788 | 73.144 |
GLCM-SVM | 92.738 | 93.714 | 93.234 | 93.412 |
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
ResNet50-KNN | 98.237 | 97.893 | 97.362 | 97.617 |
ResNet50-DT | 96.577 | 95.958 | 96.218 | 96.079 |
ResNet50-SVM | 98.340 | 98.412 | 98.178 | 98.278 |
ResNet101-KNN | 98.432 | 97.981 | 98.241 | 98.102 |
ResNet101-DT | 97.303 | 97.141 | 97.692 | 97.379 |
ResNet101-SVM | 98.651 | 98.114 | 98.235 | 98.171 |
ResNet-CBAM-KNN | 98.651 | 98.635 | 98.483 | 98.552 |
ResNet-CBAM-DT | 96.888 | 97.292 | 96.732 | 96.992 |
ResNet-CBAM-SVM | 98.444 | 98.445 | 98.441 | 98.440 |
DenseNet121-KNN | 98.548 | 98.313 | 98.313 | 98.408 |
DenseNet121-DT | 96.058 | 95.374 | 96.266 | 95.732 |
DenseNet121-SVM | 98.548 | 98.534 | 98.543 | 98.521 |
Xception-KNN | 98.029 | 97.272 | 98.289 | 97.742 |
Xception-DT | 96.680 | 96.450 | 95.893 | 96.130 |
Xception-SVM | 98.029 | 97.644 | 98.300 | 97.950 |
表5 测试集上基于卷积特征的不同分类方法比较 (%)
Tab. 5 Comparison of different classification methods based on convolutional features on test set
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
ResNet50-KNN | 98.237 | 97.893 | 97.362 | 97.617 |
ResNet50-DT | 96.577 | 95.958 | 96.218 | 96.079 |
ResNet50-SVM | 98.340 | 98.412 | 98.178 | 98.278 |
ResNet101-KNN | 98.432 | 97.981 | 98.241 | 98.102 |
ResNet101-DT | 97.303 | 97.141 | 97.692 | 97.379 |
ResNet101-SVM | 98.651 | 98.114 | 98.235 | 98.171 |
ResNet-CBAM-KNN | 98.651 | 98.635 | 98.483 | 98.552 |
ResNet-CBAM-DT | 96.888 | 97.292 | 96.732 | 96.992 |
ResNet-CBAM-SVM | 98.444 | 98.445 | 98.441 | 98.440 |
DenseNet121-KNN | 98.548 | 98.313 | 98.313 | 98.408 |
DenseNet121-DT | 96.058 | 95.374 | 96.266 | 95.732 |
DenseNet121-SVM | 98.548 | 98.534 | 98.543 | 98.521 |
Xception-KNN | 98.029 | 97.272 | 98.289 | 97.742 |
Xception-DT | 96.680 | 96.450 | 95.893 | 96.130 |
Xception-SVM | 98.029 | 97.644 | 98.300 | 97.950 |
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
GLCM+ResNet50-KNN | 98.237 | 97.760 | 97.706 | 97.732 |
GLCM+ResNet50-DT | 97.822 | 97.495 | 97.538 | 97.506 |
GLCM+ResNet50-SVM | 98.859 | 98.703 | 98.615 | 98.655 |
GLCM+ResNet101-KNN | 98.444 | 98.449 | 98.513 | 98.475 |
GLCM+ResNet101-DT | 96.058 | 95.426 | 95.823 | 95.580 |
GLCM+ResNet101-SVM | 98.859 | 98.819 | 98.619 | 98.712 |
GLCM+ResNet-CBAM-KNN | 98.652 | 98.055 | 98.853 | 98.441 |
GLCM+ResNet-CBAM-DT | 96.992 | 97.401 | 97.005 | 97.182 |
GLCM+ResNet-CBAM-SVM(本文方法) | 98.963 | 99.077 | 98.927 | 98.996 |
GLCM+DenseNet121-KNN | 98.652 | 98.794 | 98.637 | 98.700 |
GLCM+DenseNet121-DT | 95.954 | 96.498 | 96.394 | 96.361 |
GLCM+DenseNet121-SVM | 98.963 | 98.832 | 98.895 | 98.856 |
GLCM+Xception-KNN | 97.925 | 97.178 | 98.198 | 97.650 |
GLCM+Xception-DT | 97.199 | 96.691 | 97.291 | 96.979 |
GLCM+Xception-SVM | 98.029 | 97.644 | 98.300 | 97.950 |
表6 测试集上基于聚合特征的不同分类方法性能比较 ( %)
Tab. 6 Performance comparison of different classification methods based on aggregated features on test set
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
GLCM+ResNet50-KNN | 98.237 | 97.760 | 97.706 | 97.732 |
GLCM+ResNet50-DT | 97.822 | 97.495 | 97.538 | 97.506 |
GLCM+ResNet50-SVM | 98.859 | 98.703 | 98.615 | 98.655 |
GLCM+ResNet101-KNN | 98.444 | 98.449 | 98.513 | 98.475 |
GLCM+ResNet101-DT | 96.058 | 95.426 | 95.823 | 95.580 |
GLCM+ResNet101-SVM | 98.859 | 98.819 | 98.619 | 98.712 |
GLCM+ResNet-CBAM-KNN | 98.652 | 98.055 | 98.853 | 98.441 |
GLCM+ResNet-CBAM-DT | 96.992 | 97.401 | 97.005 | 97.182 |
GLCM+ResNet-CBAM-SVM(本文方法) | 98.963 | 99.077 | 98.927 | 98.996 |
GLCM+DenseNet121-KNN | 98.652 | 98.794 | 98.637 | 98.700 |
GLCM+DenseNet121-DT | 95.954 | 96.498 | 96.394 | 96.361 |
GLCM+DenseNet121-SVM | 98.963 | 98.832 | 98.895 | 98.856 |
GLCM+Xception-KNN | 97.925 | 97.178 | 98.198 | 97.650 |
GLCM+Xception-DT | 97.199 | 96.691 | 97.291 | 96.979 |
GLCM+Xception-SVM | 98.029 | 97.644 | 98.300 | 97.950 |
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
冻结全部卷积层权值 | 94.605 | 94.925 | 93.875 | 94.397 |
微调部分卷积层权值 | 94.605 | 95.175 | 93.925 | 94.545 |
表7 测试集上文献[11]中两种方法的性能比较 ( %)
Tab. 7 Performance comparison of two methods in literature [11] on test set
方法 | accuracy | macro-P | macro-R | macro-F1 |
---|---|---|---|---|
冻结全部卷积层权值 | 94.605 | 94.925 | 93.875 | 94.397 |
微调部分卷积层权值 | 94.605 | 95.175 | 93.925 | 94.545 |
1 | 中国机械工程学会热处理学会. 热处理手册:第四卷 热处理质量控制和检验[M]. 4版. 北京:机械工业出版社, 2008: 126-266. |
Chinese Heat Treatment Society. Heat Treatment Handbook: Volume IV Quality Control and Inspection of Heat Treatment[M]. 4th ed. Beijing: China Machine Press, 2008: 126-266. | |
2 | 张宏伟. 基于图像分析的废杂铜熔炼过程控制[D]. 杭州:浙江大学, 2014:2-19. |
ZHANG H W. Image analysis-based control system of scrap copper smelting processes[D]. Hangzhou: Zhejiang University, 2014:2-19. | |
3 | KIM C W, KIM H G, SUK H G. A study on the composition determination of Cu alloys by image processing technology[J]. Solid State Phenomena, 2006, 116/117: 795-798. 10.4028/www.scientific.net/ssp.116-117.795 |
4 | ZHANG H W, GE Z Q, YUAN X F, et al. Rapid vision-based system for secondary copper content estimation[J]. Transactions of Nonferrous Metals Society of China, 2014, 24(8): 2665-2676. 10.1016/s1003-6326(14)63397-5 |
5 | LI H, YAN W J, LI Z X, et al. Numerical and experimental study of the hot cracking phenomena in 6061/7075 dissimilar aluminum alloy resistance spot welding[J]. Journal of Manufacturing Processes, 2022, 77: 794-808. 10.1016/j.jmapro.2022.04.001 |
6 | 丛明,卢长奇,刘冬,等. 基于Refine-ACTDD的铸件外观微小缺陷检测方法[J]. 计算机集成制造系统, 2022, 28(9):2815-2824. |
CONG M, LU C Q, LIU D, et al. Detection method for tiny defects in casting appearance based on Refine-ACTDD[J]. Computer Integrated Manufacturing Systems, 2022, 28(9):2815-2824. | |
7 | CHEN Q L, LI W, CHEN Z. Analysis of microstructure characteristics of high sulfur steel based on computer image processing technology[J]. Results in Physics, 2019, 12: 392-397. 10.1016/j.rinp.2018.10.037 |
8 | WANG M, YANG L, ZHAO Z B, et al. Intelligent prediction of wear location and mechanism using image identification based on improved Faster R-CNN model[J]. Tribology International, 2022, 169: No.107466. 10.1016/j.triboint.2022.107466 |
9 | LI Y F, LI S H. Deep learning based phase transformation model for the prediction of microstructure and mechanical properties of hot-stamped parts[J]. International Journal of Mechanical Sciences, 2022, 220: No.107134. 10.1016/j.ijmecsci.2022.107134 |
10 | FERNANDEZ-ZELAIA P, ROSSY A M, CAMPBELL Q, et al. Austenitic parent grain reconstruction in martensitic steel using deep learning[J]. Materials Characterization, 2022, 185: No.111759. 10.1016/j.matchar.2022.111759 |
11 | PAPA J P, NAKAMURA R Y M, DE ALBUQUERQUE V H C, et al. Computer techniques towards the automatic characterization of graphite particles in metallographic images of industrial materials[J]. Expert Systems with Applications, 2013, 40(2): 590-597. 10.1016/j.eswa.2012.07.062 |
12 | DeCOST B L, HOLM E A. A computer vision approach for automated analysis and classification of microstructural image data[J]. Computational Materials Science, 2015, 110: 126-133. 10.1016/j.commatsci.2015.08.011 |
13 | 张永志,辛全忠,王永亮,等. 基于迁移学习的钢金相组织分类与识别方法的研究[J]. 材料导报, 2021, 35(24): 24152-24157. 10.11896/cldb.20070022 |
ZHANG Y Z, XIN Q Z, WANG Y L, et al. Research on classification and recognition method of steel metallographic structure based on transfer learning[J]. Materials Reports, 2021, 35(24): 24152-24157. 10.11896/cldb.20070022 | |
14 | AZIMI S M, BRITZ D, ENGSTLER M, et al. Advanced steel microstructural classification by deep learning methods[J]. Scientific Reports, 2018, 8: No.2128. 10.1038/s41598-018-20037-5 |
15 | HUMEAU-HEURTIER A. Texture feature extraction methods: a survey[J]. IEEE Access, 2019, 7: 8975-9000. 10.1109/access.2018.2890743 |
16 | BRYNOLFSSON P, NILSSON D, TORHEIM T, et al. Haralick texture features from Apparent Diffusion Coefficient (ADC) MRI images depend on imaging and pre-processing parameters[J]. Scientific Reports, 2017, 7: No.4041. 10.1038/s41598-017-04151-4 |
17 | WEBEL J, GOLA J, BRITZ D, et al. A new analysis approach based on Haralick texture features for the characterization of microstructure on the example of low-alloy steels[J]. Materials Characterization, 2018, 144: 584-596. 10.1016/j.matchar.2018.08.009 |
18 | LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 936-944. 10.1109/cvpr.2017.106 |
19 | 王志伟,普园媛,王鑫,等. 基于多特征融合的多尺度服装图像精准化检索[J]. 计算机学报, 2020, 43(4): 740-754. 10.11897/SP.J.1016.2020.00740 |
WANG Z W, PU Y Y, WANG X, et al. Accurate retrieval of multi-scale clothing images based on multi-feature fusion[J]. Chinese Journal of Computers, 2020, 43(4): 740-754. 10.11897/SP.J.1016.2020.00740 | |
20 | HARALICK R M, SHANMUGAM K, DINSTEIN I, et al. Textural features for image classification[J]. IEEE Transactions on Systems, Man, and Cybernetics, 1973, SMC-3(6): 610-621. 10.1109/tsmc.1973.4309314 |
21 | WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]// Proceedings of the 2018 European Conference on Computer Vision, LNCS 11211. Cham: Springer, 2018: 3-19. |
[1] | 刘源泂, 何茂征, 黄益斌, 钱程. 基于ResNet50和改进注意力机制的船舶识别模型[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1935-1941. |
[2] | 郭琳, 刘坤虎, 马晨阳, 来佑雪, 徐映芬. 基于感受野扩展残差注意力网络的图像超分辨率重建[J]. 《计算机应用》唯一官方网站, 2024, 44(5): 1579-1587. |
[3] | 孙敏, 成倩, 丁希宁. 基于CBAM-CGRU-SVM的Android恶意软件检测方法[J]. 《计算机应用》唯一官方网站, 2024, 44(5): 1539-1545. |
[4] | 付顺旺, 陈茜, 李智, 王国美, 卢妤. 用于篡改图像检测和定位的双通道渐进式特征过滤网络[J]. 《计算机应用》唯一官方网站, 2024, 44(4): 1303-1309. |
[5] | 张鹏飞, 韩李涛, 冯恒健, 李洪梅. 基于注意力机制和全局特征优化的点云语义分割[J]. 《计算机应用》唯一官方网站, 2024, 44(4): 1086-1092. |
[6] | 周景贤, 李希娜. 基于改进卷积神经网络和射频指纹的无人机检测与识别[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 876-882. |
[7] | 付可意, 王高才, 邬满. 基于改进区域提议网络和特征聚合小样本目标检测方法[J]. 《计算机应用》唯一官方网站, 2024, 44(12): 3790-3797. |
[8] | 乔恩保, 高向阳, 程俊. 基于支持向量机的自恢复自适应蒙特卡洛定位算法[J]. 《计算机应用》唯一官方网站, 2024, 44(10): 3246-3251. |
[9] | 张慧斌, 冯丽萍, 郝耀军, 王一宁. 基于注意力机制和迁移学习的古壁画朝代识别[J]. 《计算机应用》唯一官方网站, 2023, 43(6): 1826-1832. |
[10] | 申利华, 李波. 基于特征金字塔网络和密集网络的肺部CT图像超分辨率重建[J]. 《计算机应用》唯一官方网站, 2023, 43(5): 1612-1619. |
[11] | 蒋瑞林, 覃仁超. 基于深度可分离卷积的多神经网络恶意代码检测模型[J]. 《计算机应用》唯一官方网站, 2023, 43(5): 1527-1533. |
[12] | 宋其洪, 刘建勋, 扈海泽, 张祥平. 基于协同融合网络的代码搜索模型[J]. 《计算机应用》唯一官方网站, 2023, 43(12): 3896-3902. |
[13] | 赵欣, 祝倩倩, 赵聪, 吴佳玲. 基于多尺度和跨空间融合的超声乳腺结节分割[J]. 《计算机应用》唯一官方网站, 2023, 43(11): 3599-3606. |
[14] | 温祥西, 彭娅婷, 毕可心, 衡宇铭, 吴明功. 基于最优样本集在线模糊最小二乘支持向量机的飞行冲突网络态势预测[J]. 《计算机应用》唯一官方网站, 2023, 43(11): 3632-3640. |
[15] | 张志昂, 廖光忠. 基于U-Net的多尺度特征增强视网膜血管分割算法[J]. 《计算机应用》唯一官方网站, 2023, 43(10): 3275-3281. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||