【摘要】 砂岩薄片鉴定是矿物学和采矿工程中的一个重要步骤,其基础是将砂岩薄片图像包含的矿物颗粒分割到独立区域.不同于一般图像分割问题,砂岩薄片图像中包含大量矿物颗粒,且相邻颗粒之间边界模糊,通用的图像分割方法难以适用.本文利用多角度砂岩薄片图像,使用卷积神经网络和模糊聚类技术,提出一种3阶段颗粒分割方法.第1阶段,将输入的多角度砂岩图像预分割成超像素集合.第2阶段,根据砂岩矿物特点构建卷积神经网络RockNet,先使用带标签的砂岩矿物颗粒图像库训练RockNet,然后将之用于提取超像素语义特征.第3阶段,提出区域合并方法 FCoG,该方法融合多特征用于聚类和合并超像素,并生成最终的矿物颗粒.对采集自多个地区和不同地质年代的砂岩薄片图像数据集进行实验,结果表明本文方法的有效性,其性能明显优于其他分割方法.更多还原

【Abstract】 The identification of sandstone thin sections is a primary step in mineralogy and mining engineering,in which the principle is to partition mineral grains contained in the sandstone thin section images into separate regions. Unlike segmentation of ordinary images, thin section images of sandstone contain many mineral grains and the differences among adjacent grains are usually ambiguous, which makes conventional segmentation methods difficult to apply. In this paper, we take advantage of multi-angle thin section images of sandstone and propose a three-stage method for grain segmentation using convolutional neural networks(CNN) and fuzzy clustering. In the first stage, the input multi-angle images are pre-segmented into a set of superpixels. In the second stage, RockNet,which is a CNN based on sandstone mineral characteristics, is trained using a labeled dataset of sandstone mineral grain images and then applied to extract the semantic features of the superpixels. In the third stage, a region merging algorithm named FCoG is proposed to combine multiple features to cluster and merge superpixels to yield the final mineral grains. The experimental results conducted on the dataset of sandstone thin section images collected from multiple locations and geologic eras demonstrate the proposed method’s effectiveness, showing that it evidently outperforms other available segmentation methods.更多还原