基于深度学习特征和非线性支持向量机的板材表面缺陷识别方法
|
摘要
深度学习是一种有效的特征学习方法,具有很强的自主学习能力。研究了基于深度学习特征与非线性支持向量机(NSVM)分类算法相结合的板材表面缺陷识别方法。首先,针对深度学习模型需要海量训练数据的特性,使用旋转剪切的方法对采集到的原始板材表面缺陷图像进行数据扩增;其次,使用扩增后的板材表面缺陷图像数据集对笔者提出的深度卷积神经网络(CNN)模型进行训练,并使用训练好的网络提取不同种类缺陷图像的深度特征;然后,为了消除深度特征中的冗余数据,并增强数据的表达能力,运用基于1范数的非贪婪主成分分析(Non-greedy PCA-L1)算法对板材的深层语义特征进行特征降维和特征增强;最后,运用增强后的深度特征训练NSVM模型,并使用训练好的NSVM模型对原始板材表面缺陷图像的测试集进行分类。实验结果表明,笔者提出的识别方法具有较好的鲁棒性和实用性,可取得目前较好的分类效果,针对结疤、压痕和无损3种板材表面缺陷识别率可达99%以上。
Deep Learning( DL) is one of the most efficient feature learning methods in the field of computer vision,which has a strong autonomous learning ability. As one of the most representative DL models,the Convolutional Neural Network( CNN) model plays a fundamental role in the image processing domain. Many state-of-the-art image processing models are developed based on the CNN model. However,the CNN model usually polarizes the output values,and some of which are close to +1,while the other parts are close to-1. In practice,it is more inclined to output the attribution of certain samples in the form of a more extensive probability. In order to improve the generalization of the recognition model and make the recognition model more reliable,in this paper,the board surface defects recognition problem was investigated based on the combination of the deep learning features and the Non-linear Support Vector Machine( NSVM) classifier. Specifically,three kinds of common surface defect images were firstly collected,which were nodule images,indentation images and nondestructive images. Then,three steps were proposed for the board surface defects recognition: the collected board images were used to train a novel modified CNN classification model,and the trained CNN model was used to extract deep features of these three different surface defect images in the validation set. Then,in order to enhance the expressive ability of these depth features,the Non-Greedy Principal Component Analysis with L1-norm( Non-greedy PCA-L1) algorithm was used to deduce their dimension and obtain some more compact deep features. After that,the enhanced features were used to train the NSVM classifier. Finally,the trained NSVM classifier was used to determine the images' category with board surface defects.Experiment results demonstrated that the improved CNN model was robust and practical,and a new the-state-of-art recognition accuracy can be achieved. For these three kinds of surface defects,the recognition rate was up to 99%. In the future,it is expected to use some advanced algorithms to recognition defects on the board surface,such as the semi-supervised method. In addition,the training of CNN model requires a large number of labeled images,so the recognition method based on generation networks is also a promising research direction. The main source code and data involved in this paper will be published on the GitHub after the paper is published.
Deep Learning( DL) is one of the most efficient feature learning methods in the field of computer vision,which has a strong autonomous learning ability. As one of the most representative DL models,the Convolutional Neural Network( CNN) model plays a fundamental role in the image processing domain. Many state-of-the-art image processing models are developed based on the CNN model. However,the CNN model usually polarizes the output values,and some of which are close to +1,while the other parts are close to-1. In practice,it is more inclined to output the attribution of certain samples in the form of a more extensive probability. In order to improve the generalization of the recognition model and make the recognition model more reliable,in this paper,the board surface defects recognition problem was investigated based on the combination of the deep learning features and the Non-linear Support Vector Machine( NSVM) classifier. Specifically,three kinds of common surface defect images were firstly collected,which were nodule images,indentation images and nondestructive images. Then,three steps were proposed for the board surface defects recognition: the collected board images were used to train a novel modified CNN classification model,and the trained CNN model was used to extract deep features of these three different surface defect images in the validation set. Then,in order to enhance the expressive ability of these depth features,the Non-Greedy Principal Component Analysis with L1-norm( Non-greedy PCA-L1) algorithm was used to deduce their dimension and obtain some more compact deep features. After that,the enhanced features were used to train the NSVM classifier. Finally,the trained NSVM classifier was used to determine the images' category with board surface defects.Experiment results demonstrated that the improved CNN model was robust and practical,and a new the-state-of-art recognition accuracy can be achieved. For these three kinds of surface defects,the recognition rate was up to 99%. In the future,it is expected to use some advanced algorithms to recognition defects on the board surface,such as the semi-supervised method. In addition,the training of CNN model requires a large number of labeled images,so the recognition method based on generation networks is also a promising research direction. The main source code and data involved in this paper will be published on the GitHub after the paper is published.
引文
[1]HASHIM U R,HASHIM S Z,MUDA A K. Automated vision inspection of timber surface defect:a review[J]. Jurnal Teknologi,2015,77(20):127-135. DOI:10.11113/jt. v77.6562.
[2]HITTAWE M M,MUDDAMSETTY S M,SIDIBD,et al. Multiple features extraction for timber defects detection and classification using SVM[C]//2015 IEEE International Conference on Image Processing. IEEE,2015:427-431. DOI:10. 1109/icip.2015.7350834.
[3]DELGADO A,PEREIRA C,de BRITO J,et al. Defect characterization,diagnosis and repair of wood flooring based on a field survey[J]. Materiales de Construcción,2018,68(329):149.DOI:10.3989/mc.2018.01817.
[4]吴东洋,业宁,徐波,等.基于改进的Affnity Propagation聚类的木材缺陷识别[J].工程数学学报,2012,29(4):600-606.10.3969/j.issn. DOI:1005-3085.2012.04.017.WU D Y,YE N,XU B,et al. Wood defect recognition based on improved Affnity Propagation clustering algorithm[J]. Journal of Engineering Mathematics,2012,29(4):600-606.
[5]业宁,王厚立,徐兆军,等.基于支持向量机的木材缺陷识别[J].计算机应用与软件,2006(4):3-5. DOI:10.3969/j.issn.1000-386X.2006.04.002.YE N,WANG H L,XU Z J,et al. Wood defect recognition based on support vector machine[J]. Computer Applications and Software,2006(4):3-5.
[6]张召,业宁,业巧林.基于纹理提取和SVM技术的自动木材缺陷识别[J].计算机工程与应用,2009,45(23):219-223.DOI:10.3778/j.issn.1002-8331.2009.23.063ZHANG Z,YE N,YE Q L. Automatic wood defect recognition based on texture extraction and SVM technology[J]. Computer Engineering and Application,2009,45(23):219-223.
[7]LI C,ZHANG Y,TU W,et al. Soft measurement of wood defects based on LDA feature fusion and compressed sensor images[J].Journal of Forestry Research,2017,28(6):1285-1292. DOI:10.1007/s11676-017-0395-6.
[8]马旭,刘应安,业宁,等.基于核PCA与SVM算法的木材缺陷识别[J].常州大学学报(自然科学版),2017,29(3):60-68. DOI:10.3969/j.issn.2095-0411.2017.03.009.MA X,LIU Y A,YE N,et al. Wood defect identification based on nuclear PCA and SVM algorithm[J]. Journal of Changzhou University(Natural Science Edition),2017,29(3):60-68.
[9]谢永华,陈庆为,梁娇娇.基于改进K-means聚类的木材缺陷彩色图像分割算法研究[J].现代科学仪器,2014(3).DOI:10.7666/d. Y2234439.XIE Y H,CHEN Q W,LIANG J J. Research of wood defect color image segmentation algorithm based on improved K-means clustering algorithm[J]. Modern Science Instrument,2014(3).
[10]徐姗姗,刘应安,徐昇.基于卷积神经网络的木材缺陷识别[J].山东大学学报(工学版),2013,43(2):23-28. DOI:10.6040/j.issn.1672-3961.1.2012.213.XU S S,LIU Y A,XU S. Wood defect recognition based on convolution neural network[J]. Journal of Shandong University(Engineering Edition),2013,43(2):23-28.
[11]REN R X,HUNG T,TAN K C. A generic deep-learning-based approach for automated surface inspection[J]. IEEE Transactions on Cybernetics,2018,48(3):929-940. DOI:10. 1109/tcyb.2017.2668395.
[12]ZHANG D,SUN Y,YE Q,et al. Recursive discriminative subspace learning with L1-Norm distance constraint[J]. IEEE Transactions on Cybernetics,2018,1(1):1-14. DOI:10. 1109/tcyb.2018.2882924.
[13] KRIZHEVSKY A,SUTSKEVER I,HINTON G E. Image-net classification with deep convolutional neural networks[C]//International Conference on Neural Information Processing Systems.Curran Associates Inc. 2012:1097-1105. DOI:10. 1109/ijcnn.2018.8489274.
[14]DING J,CHEN B,LIU H,et al. Convolutional neural network with data augmentation for SAR target recognition[J]. IEEE Geoscience&Remote Sensing Letters,2016,13(3):364-368.DOI:10.1109/lgrs.2015.2513754.
[15]JIN X J,XU C,FENG J,et al. Deep learning with S-shaped rectified linear activation units[J]. National Conference on Artificial Intelligence, 2016:1737-1743. DOI:10. 1109/lgrs. 2015.2513754.
[16]KWAK N. Principal component analysis based on L1-norm maximization[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2008,30(9):1672-1680. DOI:10. 1109/tpami.2008.114.
[17]NIEF P,HUANG H,DING C H,et al. Robust principal component analysis with non-greedy l 1-norm maximization[C]//International Joint Conference on Artificial Intelligence. 2011:1433-1438. DOI:10.1016/j.neucom.2015.06.011.
[18]ZHOU C J,WEI X,ZHANG Q,et al. Fisher’s linear discriminant(FLD)and support vector machine(SVM)in non-negative matrix factorization(NMF)residual space for face recognition[J]. Optica Applicata,2010,40(3):693-704. DOI:10.1109/ifcsta.2009.334.
[19]DONG S J,SUN D,TANG B,et al. A fault diagnosis method for rotating machinery based on PCA and Morlet Kernel SVM[J].Mathematical Problems in Engineering,2014:1-8. DOI:10.1155/2014/293878.
[20]JIA Y Q,SHELHAMER E,DONAHUE J,et al. Caffe:convolutional architecture for fast feature embedding[J]. ACM Multimedia,2014:675-678. DOI:10.1145/2647868.2654889.
[21] MOOSAVI-DEZFOOLI S M, FAWZI A, FROSSARD P.Deepfool:a simple and accurate method to fool deep neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016:2574-2582. DOI:10.1109/cvpr.2016.282.
[2]HITTAWE M M,MUDDAMSETTY S M,SIDIBD,et al. Multiple features extraction for timber defects detection and classification using SVM[C]//2015 IEEE International Conference on Image Processing. IEEE,2015:427-431. DOI:10. 1109/icip.2015.7350834.
[3]DELGADO A,PEREIRA C,de BRITO J,et al. Defect characterization,diagnosis and repair of wood flooring based on a field survey[J]. Materiales de Construcción,2018,68(329):149.DOI:10.3989/mc.2018.01817.
[4]吴东洋,业宁,徐波,等.基于改进的Affnity Propagation聚类的木材缺陷识别[J].工程数学学报,2012,29(4):600-606.10.3969/j.issn. DOI:1005-3085.2012.04.017.WU D Y,YE N,XU B,et al. Wood defect recognition based on improved Affnity Propagation clustering algorithm[J]. Journal of Engineering Mathematics,2012,29(4):600-606.
[5]业宁,王厚立,徐兆军,等.基于支持向量机的木材缺陷识别[J].计算机应用与软件,2006(4):3-5. DOI:10.3969/j.issn.1000-386X.2006.04.002.YE N,WANG H L,XU Z J,et al. Wood defect recognition based on support vector machine[J]. Computer Applications and Software,2006(4):3-5.
[6]张召,业宁,业巧林.基于纹理提取和SVM技术的自动木材缺陷识别[J].计算机工程与应用,2009,45(23):219-223.DOI:10.3778/j.issn.1002-8331.2009.23.063ZHANG Z,YE N,YE Q L. Automatic wood defect recognition based on texture extraction and SVM technology[J]. Computer Engineering and Application,2009,45(23):219-223.
[7]LI C,ZHANG Y,TU W,et al. Soft measurement of wood defects based on LDA feature fusion and compressed sensor images[J].Journal of Forestry Research,2017,28(6):1285-1292. DOI:10.1007/s11676-017-0395-6.
[8]马旭,刘应安,业宁,等.基于核PCA与SVM算法的木材缺陷识别[J].常州大学学报(自然科学版),2017,29(3):60-68. DOI:10.3969/j.issn.2095-0411.2017.03.009.MA X,LIU Y A,YE N,et al. Wood defect identification based on nuclear PCA and SVM algorithm[J]. Journal of Changzhou University(Natural Science Edition),2017,29(3):60-68.
[9]谢永华,陈庆为,梁娇娇.基于改进K-means聚类的木材缺陷彩色图像分割算法研究[J].现代科学仪器,2014(3).DOI:10.7666/d. Y2234439.XIE Y H,CHEN Q W,LIANG J J. Research of wood defect color image segmentation algorithm based on improved K-means clustering algorithm[J]. Modern Science Instrument,2014(3).
[10]徐姗姗,刘应安,徐昇.基于卷积神经网络的木材缺陷识别[J].山东大学学报(工学版),2013,43(2):23-28. DOI:10.6040/j.issn.1672-3961.1.2012.213.XU S S,LIU Y A,XU S. Wood defect recognition based on convolution neural network[J]. Journal of Shandong University(Engineering Edition),2013,43(2):23-28.
[11]REN R X,HUNG T,TAN K C. A generic deep-learning-based approach for automated surface inspection[J]. IEEE Transactions on Cybernetics,2018,48(3):929-940. DOI:10. 1109/tcyb.2017.2668395.
[12]ZHANG D,SUN Y,YE Q,et al. Recursive discriminative subspace learning with L1-Norm distance constraint[J]. IEEE Transactions on Cybernetics,2018,1(1):1-14. DOI:10. 1109/tcyb.2018.2882924.
[13] KRIZHEVSKY A,SUTSKEVER I,HINTON G E. Image-net classification with deep convolutional neural networks[C]//International Conference on Neural Information Processing Systems.Curran Associates Inc. 2012:1097-1105. DOI:10. 1109/ijcnn.2018.8489274.
[14]DING J,CHEN B,LIU H,et al. Convolutional neural network with data augmentation for SAR target recognition[J]. IEEE Geoscience&Remote Sensing Letters,2016,13(3):364-368.DOI:10.1109/lgrs.2015.2513754.
[15]JIN X J,XU C,FENG J,et al. Deep learning with S-shaped rectified linear activation units[J]. National Conference on Artificial Intelligence, 2016:1737-1743. DOI:10. 1109/lgrs. 2015.2513754.
[16]KWAK N. Principal component analysis based on L1-norm maximization[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2008,30(9):1672-1680. DOI:10. 1109/tpami.2008.114.
[17]NIEF P,HUANG H,DING C H,et al. Robust principal component analysis with non-greedy l 1-norm maximization[C]//International Joint Conference on Artificial Intelligence. 2011:1433-1438. DOI:10.1016/j.neucom.2015.06.011.
[18]ZHOU C J,WEI X,ZHANG Q,et al. Fisher’s linear discriminant(FLD)and support vector machine(SVM)in non-negative matrix factorization(NMF)residual space for face recognition[J]. Optica Applicata,2010,40(3):693-704. DOI:10.1109/ifcsta.2009.334.
[19]DONG S J,SUN D,TANG B,et al. A fault diagnosis method for rotating machinery based on PCA and Morlet Kernel SVM[J].Mathematical Problems in Engineering,2014:1-8. DOI:10.1155/2014/293878.
[20]JIA Y Q,SHELHAMER E,DONAHUE J,et al. Caffe:convolutional architecture for fast feature embedding[J]. ACM Multimedia,2014:675-678. DOI:10.1145/2647868.2654889.
[21] MOOSAVI-DEZFOOLI S M, FAWZI A, FROSSARD P.Deepfool:a simple and accurate method to fool deep neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016:2574-2582. DOI:10.1109/cvpr.2016.282.