Introduction
Microplastic pollution represents a significant environmental challenge, yet its quantification remains a manual, labor-intensive bottleneck. Researchers at ETH Zurich have addressed this by developing an automated computer vision pipeline designed specifically for the high-throughput analysis of microplastic fibers. Unlike standard objects, these fibers are thin, elongated, and frequently overlap, creating "tangled" clusters that defeat general-purpose instance segmentation models like Mask R-CNN.
Methodology: A Hybrid Pipeline
The proposed approach moves away from pure end-to-end deep learning, instead utilizing a multi-stage hybrid pipeline that combines traditional image processing with targeted deep learning.
The process begins with locally adaptive thresholding to extract a binary mask from grayscale microscopy images. This mask is then reduced to a one-pixel-wide morphological skeleton, identifying the fundamental structure and junctions of the fibers.
The core innovation lies in resolving these junctions. While simple intersections are handled via geometric continuity rules, complex overlaps are processed using Deep Pixel Embeddings (DPE). This model maps pixels into a high-dimensional vector space where those belonging to the same fiber instance are clustered together, allowing the system to accurately disentangle overlapping branches based on local context.
Performance and Metrics
To evaluate the system, the researchers introduced a modified Intersection-over-Union (mIoU) metric. Standard IoU is overly sensitive to the minor pixel-level misalignments common with thin structures; mIoU uses binary dilation to provide a more robust measure of segmentation quality.
On a dataset provided by EMPA containing thousands of manually annotated fibers, the DPE-enhanced pipeline significantly outperformed baselines. In "tangled" scenarios, it achieved a precision and recall of approximately 0.41 and 0.43 respectively, nearly doubling the performance of Mask R-CNN (0.23). Furthermore, the system's counting accuracy (within 7 ± 2% of ground truth) approaches the level of consistency found between human experts (4 ± 1%).
Conclusion
By decomposing the segmentation task into geometric skeletonization and local embedding-based disentanglement, this research provides a robust tool for environmental scientists. This automated approach not only saves time but also mitigates the systematic bias where manual counters often overlook or mismeasure complex, tangled fibers—which are typically 63% longer than isolated ones. Future work will focus on improving the handling of translucent fibers and expanding the dataset diversity.
Summary updated