Mikroplastikų segmentacija ir klasifikavimas naudojant neuroninius tinklus / ; Segmentation and classification of microplastics using neural networks.

This bachelor’s thesis addresses the problem of automated microplastic analysis using neural network methods. The aim is to develop an integrated system that combines microplastic segmentation with unsupervised learning classification based on morphological features. The research utilizes a dataset of microscopic microplastic images with corresponding segmentation masks. The developed system operates in four stages: first, a UNet architecture neural network model performs microplastic segmentation on nonhomogeneous backgrounds; the second stage extracts morphological features (circularity, elongation ratio, solidity) from segmented objects; the third stage employs Kmeans clustering to group microplastics into three main morphological categories; the fourth stage generates pseudolabels and uses a MobileNetV3Small convolutional neural network for final classification, integrating image features with morphological parameters. The UNet model achieved a 0.75 Dice coefficient in the segmentation task, while the hybrid MobileNetV3Small classification model reached 68.2% accuracy for categorizing microplastics into fragments, granules, and fibers. The developed system demonstrated suitability for practical automation of microplastic analysis, although it requires further improvement with larger datasets.