Microplastics (MPs) are an emerging pollutant of global concern, creating an urgent need for rapid and accurate monitoring workflows. Deep learning-based computer vision has demonstrated strong performance in finding particles in microscopy images, but its use as a front-end module for automated IR/Raman microscope-based MP analysis remains insufficiently developed, particularly in workflows that convert image-level particle detection results into microscope-executable operations for particle morphological characterization and spectral acquisition planning. Here, we aim to address this gap. First, reference MPs were deposited on Anodisc filters and glass slides, from which 626 bright-field micrographs containing 7010 particles were collected. The dataset was randomly split into training, validation, and test sets in a 7 : 2 : 1 ratio, and an Ultralytics YOLO11 instance segmentation model was developed. At an intersection-over-union (IoU) threshold of 0.7, testing achieved precision, recall, and F1 scores of 0.86, 0.93, and 0.89, respectively, outperforming two benchmark methods: a threshold-based method (F1 = 0.25) and Mask R-CNN (F1 = 0.84). On another test set prepared from tap water (with a relatively clean background), the model achieved precision, recall, and F1 scores of 0.78, 0.90, and 0.83 at IoU = 0.7. However, particle detection performance decreased with dirtier image backgrounds, as shown by testing on the sample prepared from commercial salt. We addressed key technical challenges required for end-to-end automation, including per-field autofocus, recovering true microscope coordinates from YOLO style outputs, converting detections into particle descriptors for downstream characterization, and programmatic microscope control for image collection and spectrum acquisition planning. Full implementation, code, and well annotated data are released openly, enabling adoption and extension of this workflow for broader MP monitoring applications.
更多
原文链接
分享
引用
