Identifying Microparticle Clusters in High Resolution: Computer Vision Applied to Polymer Particles in Liquid Crystal (LC) to Enable On-the-Fly Characterization of Their Morphology and Size

Student’s name: Michael Batavia Home Institution: NYU Tandon School of Engineering NNCI Site: CNF @ Cornell University REU Principal Investigator: Prof. Fengqi You and Prof. Nicholas Abbott REU Mentors: Soumyamouli Pal and Guangyao Chen Abstract: In-situ monitoring of polymer particle formation by iCVD-in-LC is limited by relatively low-resolution (LR) capabilities compared to its ex-situ microscopy approaches. In this paper, we address this challenge by fine-tuning and developing an object-oriented super resolution model to improve the image resolution obtained from the in-situ monitoring and enable identification and characterization (e.g. size estimation) of individual polymer particles both singly dispersed and aggregated as clusters. This framework will transform low-resolution images into high-resolution object-level high-resolved (HR) images by learning from LR–HR image pairs obtained from samples observed both in the iCVD reactor (LR) and in the laboratory using a high-resolution microscope (HR). Current methods are constrained by the use of object detection, which is limited by the time needed to manually label the clusters during both train and test time for input into the super-resolution neural network. By contrast, our method uses classical computer vision techniques to isolate the slides where the particle clusters lie, keeping spatial awareness and allowing for scalability and little to no preprocessing for inputting LR images into the super resolution network. In this project, we used 5µm polystyrene particles dispersed in 5CB liquid crystal as a surrogate for polymerization via iCVD as samples for the LR and HR dataset. To align the image pairs to maximize the surface area of the slide and clusters shown, we converted images to grayscale, applied a Gaussian blur (σ = 20), identified and filtered the image contours with maximum area via the marching squares algorithm, and added a Hough Transform on the LR images to correct for any rotational vibrations added by the iCVD reactor during imaging. These slides were split into a 80/10/10 train/test/validation split and fine-tuned on the Real-ESRGAN super resolution model with pretrained weights. Resolution upgrade was quantified using the peak signal-to-noise ratio metric during validation and testing. In the end, we show that super resolution neural nets may present an alternative to predicting the characteristics of iCVD polymer formation compared to post ex-situ microscopy approaches with only use of the LR iCVD reactor image set.