Panoptic Segmentation of Animal Fibers

Animal fiber identification is an essential aspect of fabric production, since specialty fibers such as cashmere are often targeted by adulteration attempts. Current, automated fiber identification methods are often based on the optical analysis of fiber surface morphology, and require a panoptic segmentation (i.e. the complete, non-overlapping instance segmentation) of animal fibers. To date, these are provided manually in a labor-intensive manner, reducing the applicability of developed solutions. In our work, we tackle the automated, panoptic segmentation of animal fibers, overcoming the above limitation. We propose a two-step procedure consisting of (I) the non-overlapping, binary segmentation of fiber "scale edges", followed by (II) the rule-based conversion of the binary "scale edge" mask to the panoptic segmentation of the animal fiber. For the first step, we investigate whether semi-supervised learning outperforms supervised learning in the low-data regime. We motivate this by the fact that acquiring fiber scale images is much less time-consuming than segmenting them, and find it not to be the case. For the second step, we propose a rule-based post-processing of generated "scale edge" heatmaps for improved separability of "fiber scale" instances, and show that the post-processing improves performance across all evaluated configurations. In total, we demonstrate that the automated, panoptic segmentation of animal fibers is feasible.