【作者】 刘颖；

【导师】 陈大力；

【作者基本信息】 东北大学 ， 系统工程， 2020， 硕士

【摘要】 钢铁表面缺陷是决定钢铁产品质量的重要因素,在钢铁生产过程中,产品质量检测是必不可少的阶段。钢铁表面缺陷的自动分类是钢铁产品质量智能分析系统中的关键环节,对钢铁企业的高质量发展至关重要。本文以某大型钢铁企业的钢铁产品表面缺陷图像为对象,结合机器学习和深度学习先进理论,对基于自步学习的钢铁表面缺陷图像分类方法进行了深入研究,实现了钢铁表面缺陷图像的自动分类,主要研究工作如下:(1)针对钢铁表面缺陷图像的自动分类问题,提出了基于自步支持向量机的钢铁表面缺陷图像分类方法。该方法首先根据钢铁表面缺陷特点,构建了具有强判别性的缺陷特征池,实现了钢铁表面缺陷的准确表示。在此基础上,将目前先进的自步学习方法与支持向量机分类方法相结合,构造了自步支持向量机分类模型。针对该模型的参数学习问题,设计了一种ACS(Alternative Convex Search)方法,实现了模型参数‘由简至难’的学习过程,提高了缺陷分类的准确性。为了验证分类方法有效性,构造了钢铁表面缺陷数据集。大量的对比实验结果表明,该方法能够实现钢铁表面缺陷图像的准确分类,同时自步学习有助于提高支持向量机的分类性能。(2)针对钢铁表面缺陷图像标记工作成本高的问题,提出了基于自步学习的无监督钢铁表面缺陷聚类方法。该方法将自步学习方法与高斯混合模型相结合,提出了自步多样性高斯混合聚类模型。针对模型参数的学习问题,实现了一种高效的交替搜索策略,这种策略能够在保障‘由简至难’学习的同时,还能够保障学习的多样性要求。实验结果表明,自步多样性高斯混合聚类方法的聚类性能优于传统的高斯混合模型。(3)针对钢铁表面缺陷特征设计困难的问题,提出了基于自步Densenet的钢铁表面缺陷图像分类方法。该方法将自步学习方法与深度神经网络相结合,构造了自步Densenet分类模型。针对模型参数的学习问题,实现了一种高效的交替搜索策略,这种策略能够在自动学习缺陷特征的同时,实现模型参数‘由简至难’的学习。实验结果表明,该方法能够获得优于Densenet的分类性能,同时还具有很好的鲁棒性。更多还原

【Abstract】 Surface defects of steel are essential factors that determine the quality of steel products.In the process of steel production,product quality inspection is a crucial stage.The automatic classification of steel surface defects is a key section in the intelligent analysis system for the quality of steel products,which is essential for the high-quality development of steel companies.This paper takes the surface defect images of steel products of a large iron and steel enterprise as the object,and combines advanced theories of machine learning and deep learning to conduct an in-depth study on the classification method of steel surface defects based on self-paced learning,and realize automatic classification of steel surface defect images.The main research work is as follows:(1)Aiming at the problem of automatic classification of steel surface defect images,this paper proposes a method of steel surface defect image classification based on a self-paced supporting vector machine.According to the characteristics of steel surface defects,this method first constructs a defect feature pool with strong discriminability,and realizes the accurate representation of steel surface defects.On this basis,the current advanced self-paced learning method and support vector machine classification method are combined to construct a self-paced support vector machine classification model.Aiming at the learning problem of this model,this paper designs an ACS(Alternative Convex Search)method to realize the learning process of the classification model ’from simple to difficult’ and improve the accuracy of defect classification.In order to verify the validity of the method,a steel surface defect data set was constructed.A large number of comparative experimental results show that the method can achieve the accuracy of steel surface defect classification,and self-paced learning can help improve the classification performance of support vector machines.(2)Aiming at the problem of high cost of image marking on steel surface defects,an unsupervised clustering method for steel surface defects based on self-paced learning is proposed.This method combines a self-paced learning method with a Gaussian mixture model,and proposes a self-paced diversity Gaussian mixture clustering model.Aiming at the learning problem of model parameters,an efficient alternative search strategy has been implemented.This strategy can guarantee to learn from ’easy to difficult’ while also ensuring the diversity of learning requirements.The experimental results show that the clustering performance of the self-paced Gaussian mixture clustering method is superior to the traditional Gaussian mixture model.(3)Aiming at the difficulty in designing the surface defect features of steel,an image classification method for steel surface defects based on self-paced Densenet is proposed.This method combines a self-paced learning method with a deep neural network,and proposes a selfpaced Densenet classification model.Aiming at the learning problem of model parameters,an efficient alternative search strategy was implemented.This strategy can automatically learn the defect features and realize the learning process of model parameters from simple to difficult.Experimental results show that this method can achieve better classification performance than Densenet,and also has good robustness.更多还原