Object Detection in SEM Images Using Convolutional Neural Networks: Application on Pyrite Framboid Size-Distribution in Fine-Grained Sediments

Abstract:

Pyrite framboids (FeS₂) occur in sediments of all geological ages and are characterized by an external spheroidal form and an internal discrete, equant microcrystalline architecture. Framboid size distributions are established during early diagenesis, are preserved through advanced stages of diagenesis, and are used to infer depositional redox conditions. Determination of framboid size distributions by manual tracing, is slow, laborious, and prone to human error. In back-scattered electron images and in EDS elemental mapping pyrite framboids stand out in sharp contrast to surrounding silicate and carbonate minerals, making them an interesting test case for machine learning because of the relative ease of their discrimination from other components.

In this work, we implement deep learning techniques to characterize framboids from 14 samples across depth in the Marcellus Shale (Devonian). We leverage deep learning techniques such as convolutional neural networks (CNNs) to replace manual tracing and to reduce significantly the time for object detection and image segmentation. Specifically, we use several architectures, including Inception, Resnet, and their modified CNN architectures. A total of ~ 1800 framboids from 128 scanning electron microscope (SEM) images were characterized to delineate the changes in water column oxygenation during the deposition of the Marcellus Shale. Among 128 images, 33 were tested and 95 were used for training. The results of the CNN algorithm were compared to manually traced results using JMicroVision. Significantly, we find that the CNN algorithm can detect up to ~ 95 % of the total number of framboids traced manually. We assess the accuracy of the CNN algorithm by comparing the equivalent framboid diameter calculated by the CNN methods to those measured through manual tracing and find greater than 95% accuracy. This level of accuracy for framboids suggests that less-distinct rock components may require larger training sets to provide satisfactory accuracy.