【作者】 李娜；

【导师】 顾庆；

【作者基本信息】 南京大学 ， 计算机技术（专业学位）， 2017， 硕士

【摘要】 高效准确的砂岩鉴别对建筑选材、资源勘探、油气储集层评价等方面有非常重要的意义。传统的人工鉴别方式个人主观性强,缺乏复用性,且耗费大量时间,鉴别效率极低。随着计算机技术的发展,砂岩薄片被制备成显微图像,结合图像处理技术和机器学习方法,快速地实现砂岩薄片的自动分类。为保证预测模型的准确性和可靠性,传统的机器学习假设训练样本与测试样本独立同分布且足够可用。然而,当需要对陌生地区的砂岩进行鉴别时,这两个假设往往无法满足。其一,由于人工标注过于费时,往往缺少足够的训练数据;其二,由于形成环境以及图像采集条件的不同,不同地区砂岩薄片显微图像存在差异。针对以上跨地区砂岩薄片显微图像自动分类问题,本文提出两种解决方案:1)基于近邻法的数据集重构,2)基于迁移学习的Festra方法。前者采用近邻法为目标地区数据构建新的训练集,并进一步精化,最后基于新的训练集训练合适的目标分类器。后者结合特征选择和改进的TrAdaBoost方法训练目标分类器。两者的目的都是减小源地区和目标地区砂岩薄片显微图像的数据分布差异,提高目标地区砂岩薄片显微图像的分类准确性,进而提高砂岩鉴别效率。实验结果证明了两种方法的有效性。本文的主要贡献概括如下:1.提出基于近邻法的跨地区砂岩薄片显微图像自动分类方法。本文提出一种基于近邻法的数据集重构方法,基于欧氏距离采用近邻法为目标地区数据选择相似的源地区实例构成新的训练集,并通过特征选择和类间平衡进一步对其精化,最后基于新的训练集训练合适的目标分类器。其主要目的是减小源地区和目标地区砂岩薄片显微图像数据的分布差异,提高砂岩鉴别的效率。2.提出基于迁移学习的跨地区砂岩薄片显微图像自动分类方法。针对砂岩薄片显微图像自动分类中存在的数据分布差异性,本文提出一种基于特征迁移和实例迁移结合的Festra方法,首先通过特征选择减小源地区和目标地区特征的不匹配性,排除冗余和无用特征;再通过改进的TrAdaBoost方法进行实例迁移,减小地区间的数据分布差异,从而提高砂岩薄片显微图像自动分类的准确性。3.对Festra方法在跨地区砂岩薄片显微图像自动分类问题的有效性进行了实证研究。本文针对四个不同角度的研究问题,基于采集自西藏不同地区不同地质年代的砂岩薄片显微图像分别设计并进行了四组实验,尽可能地对Festra方法的有效性进行全面评估,包括使用不同的评价指标、不同的分类模型和不同类型的相似性度量等。实验结果表明Festra方法可以有效提高跨地区砂岩显微图像分类的预测性能。另外,本文还对方法可能的改进和扩展进行了实证研究和结果分析。更多还原

【Abstract】 Efficient identification of sandstones is significant for the selection of building ma-terials,resource exploration and reservoir evaluation of oil and gas.Traditional manual method is subjective and time-consuming,and lacks reusability.As the development of computer technology,the sandstone thin sections are taken into microscopic images and are quickly identified by the image processing and machine learning method.To ensure the accuracy and reliability of the prediction model,traditional machine learn-ing assumes that the training samples are enough and the distributions between training and testing samples are identical and independent.However,these assumptions do not hold when identifying the sandstones collected from an unfamiliar region.On one hand,the training data will be not enough as the manual labeling is time-consuming.On the other hand,due to the diverse formation environments and acquisition condi-tions,the sandstone microscopic images vary from the regions.For the above interregional sandstone microscopic image automatic classification problem,we propose two solutions:one is nearest neighbor based dataset reconstruc-tion,the other is transfer learning based method called Festra.The former reconstructs a new training set for target data by nearest neighbor algorithm and further elabora-tion,then trains an accuracy target classifier based on the new training set.The lat-ter trains a high-quality target classifier by combining feature selection and enhanced TrAdaBoost.The goal of both methods is reducing the distribution difference between source and target region,and improving the performance and efficiency on the sand-stone microscopic image classification.The experimental results show the validity of both methods.The main contributions can be summarized as follows:1.We propose a nearest neighbor based method for interregional sandstone microscopic image classification.We propose a nearest neighbor based dataset reconstruction method to reduce the distribution difference between source and target region.The method reconstructs a new training set by nearest neighbor algorithm based on the Euclidean distance,and elaborates it through feature se-lection and class balance.Finally the target classifier is trained based on the new training set.2.We propose a method for interregional sandstone microscopic image clas-sification based on transfer learning.We propose a transfer learning method Festra for the distribution difference to improve the accuracy of the sandstone microscopic images classification.Festra contains two parts:one is feature se-lection,which mainly screens out the irrelevant and redundant features,the other is enhanced TrAdaBoost,which reduces the difference between the source and target region.3.We conduct empirical studies to assert the validity of Festra for interre-gional sandstone microscopic image classification.Based on the sandstone microscopic images taken from different regions and geologic time in Tibet,we conduct experiments according to the four research questions embarked from different perspectives.We use different performance measures,classifiers and similarity measures to evaluate the validity of Festra as completely as possible.The experiment results show the effectiveness of Festra.Moreover,we also con-duct empirical studies and result analyses for the possible improvements and expansions of the method.更多还原