A Deep Learning System for Detecting Microplastics Particles

Classification and quantification of microplastics particles is a crucial task in tracking plastics contamination and understanding its impact on the environment. However, it has traditionally been a tedious and laborious task when done manually, or it required special equipment when automatized, thus raising the need for a faster and more affordable process. Previous research on automatized microplastics classification using only digital camera images relied on instance segmentation, with the premise that the shape of the particle is a cue on its origin. This might be true while classifying only three types of microplastics: lines, pellets and fragments, which differentiate significantly in their shape. However, when also considering organic particles and tar, which are oftentimes a byproduct of microplastics collection and display shapes similar to microplastics particles, the segmentation based approaches face a significant drop in performance. This work introduces a Deep Learning approach for counting and classifying microplastics particles from an image taken with a digital camera or a mobile phone with resolution of 16M pixels or higher. The suggested approach comprises two steps: image processing using the Mosaic with Overlap technique, and object detection using YOLOv5 convolutional neural network. Five experiments are done, one concerning the image processing step, and four concerning data augmentation and transfer learning. The state of the art system for automatic classification and quantification of microplastics particles not relying on special equipment was compared to the model devised in this work in terms of precision, recall and F1 metric, and the results were extremely satisfactory: YOLOv5-based model shows 303% higher Precision value, 490% increase in Recall, and 391% higher F1 value for all classes. Thus, this is the very first deep learning-based system which can suceessfully classify the three microplastics classes, as well as tar and organic particles with high Precision, Recall and F1 rates (>98%)