Automated Segmentation of Profiled Fibers in cross-sectional Micrographs for Quality Control

摘要

The industrial standard for quality control of pro-filed fibers is the visual inspection of a batch's deviation from the production master. This is performed by human experts using cross-sectional micrographs. An alternative to this qualitative analysis is to use metrics describing the desired morphological characteristics of the fiber, so-called shape factors, in a quantitative approach. Such a quantitative approach requires the manual segmentation of profiled fibers in cross-sectional micrographs, and is therefore time-consuming and economically infeasible. Thus, an automated segmentation of profiled fibers is required to facilitate industrial use of quantitative shape factor analysis. We develop and present a pipeline for automated instance segmentation of profiled fibers in cross-sectional micrographs. The pipeline consists of preprocessing, seed generation, segmentation and candidate selection steps. Different configurations of the pipeline were investigated and their performance was evaluated across four profiled fiber types using manual segmentations as groundtruth. For the best configuration, a median F1 score of 0.935 could be achieved over all four evaluated fiber types. The developed pipeline is publicly available and expected to facilitate the quantitative quality control of profiled fibers.