Deep Learning-Based Instance Segmentation to Characterize the Morphology of Compact Aggregates through Image Analysis

The morphological characterization of aggregates is crucial in various industries, affecting the physicochemical properties and functionality of materials. This study develops a dataset of 4,000 synthetic images that closely represent real latex aggregates, validated by Frechet Inception Distance (FID) computations using a stochastic geometrical model. It also compares five instance-based deep learning segmentation models across three architectures (Mask R-CNN, YOLOv8, and SAM) for analyzing the morphology of latex aggregates. Among them, Mask R-CNN with ResNet101 showed superior segmentation quality. When applied to real images taken at different stages of aggregation, the segmentation results of this model closely matched the experimental observations, demonstrating its capability for detailed morphological analysis.