基于图像分类的矿物含量测定及精度评价

Mineral contents determination and accuracy evaluation based on classification of petrographic images

针对传统矿物含量测定中存在人为误差、缺乏精度评价等问题,提出了基于图像分类的矿物含量测定及精度评价方法,该方法通过统计分类后图像中每种矿物的像元数量测定矿物含量,并采用混淆矩阵评价含量测定精度.根据岩石图像的光谱和纹理特征,提出了两种基本的矿物含量测定方式:1)对于纹理简单、矿物光谱区分度大的岩石图像,采用直接分类方式测定矿物含量,花岗岩手标本照片矿物分类实验表明监督分类效果优于非监督分类,且监督分类中最大似然法分类(MLC)的精度最高,其含量测定精度为94.25%;2)针对复杂纹理(如干涉色、双晶等)的岩石图像,引入了面向对象(矿物或矿物集合体)的多尺度图像分割算法,在分割基础上分类并统计每类矿物含量.白云母二长花岗岩镜下照片矿物分类实验得到其含量测定精度为94.85%.

There are many human errors and lack accuracy evaluation existing in the mineral contents determination for traditional methods.A new approach is proposed for mineral contents determination and accuracy evaluation based on images classification.The method is firstly to divide the petrographic images into different mineral classes by using image classification algorithms,and then to obtain the mineral contents through pixel statistic,finally contents accuracy evaluation is carried out by Confusion Matrix（CM）.According to the spectral and texture features of the petrographic images,two approaches were proposed for mineral contents determination.One is for the petrographic images with simple texture and large color distinction is to adopt direct classification.The experiment of granite photos shows that the supervised classifiers are better than the unsupervised ones in accuracy,and the Maximum Likelihood Classifier（MLC） results with the highest accuracy of 94.25%;The other method is for the petrographic images with complex mineral texture（such as interference colors,twins,etc.）,an object-oriented Multi-resolution Segmentation（MS）algorithm is employed for images segmentation before mineral classification.The muscovite monzogranite microscope image experiment shows the content estimated accuracy is 94.85%.