Technique of the identification, quantification and measurement of carbon short-fibers in SEM images using the instance segmentation

Abstract

This paper demonstrates the use of a neural network with additional training on synthetic data to identify, quantify, and measure short carbon fibers in electron microscope photographs. This task is of importance for the development of a short carbon fiber reinforced polymer material model, which requires precisely counting and measuring the fibers in a sample to determine the structural characteristics of the material. To automate the process of counting and measuring fibers, a neural network architecture called Mask R-CNN was chosen, which is designed to implement computer vision techniques such as: object identification, segmentation and quantification of instances. The selection of this type of architecture was due to the advantages of giving the masks for each instance, which allows obtaining approximate measurements of the fiber geometry. Due to the unavailability of fiber image data, the virtual imaging technique was chosen. Artificial images of short carbon fibers were recreated using the open API NX. A virtual data set with different fiber layouts was created. The results obtained in the test phase are good, for small numbers of fibers and with sparse clusters. There are still some problems in fully identifying the geometry of fibers that overlap with other fibers, which is a challenge to solve in future work.