Deep Learning for Instance Segmentation with Object Tracking for Microscopic Crystallization Kinetics of Homopolymer in Dilute Phase

Understanding the crystallization kinetics of homopolymers is important for the design and manufacturing of recycled polymers with targeted properties. Analyzing Polarized Optical Microscopy (POM) images is widely used in investigating the morphology and micro-structure of polymers, which is often done manually in experiments, time-consuming and prone to human error, besides very limited information can be accessed. We replace the classical tedious manual labor with a few-shot learning data-driven model by a combination of deep learning-based segmentation and object tracking. Specifically, we compared the performance of both computer vision algorithm (Otus’ method) and instance segmentation frameworks, such as, Mask R-CNN (Mask Region-based Convolutional Neural Network), YOLO (You ONLY LOOK ONCE), and SAM (Segment Anything Model). Then we integrate YOLO for segmentation with DeepSORT for object tracking. After that, we deploy the above framework to analyze POM images. We first input our annotated data to a pretrained YOLO model and fine tune it to develop our own downstream models. We then extract physical insights from the outputs, such as number, size and distribution of the shaded masks. We further use a tracker based on DeepSORT algorithm to track the change of shaded masks of each instance across different frames. The results show that crystallization, growth, radius distribution and growth rate of spherulites can be accurately captured by the above segmentation framework.