Oil particle analysis using machine learning and holography imaging

It is advantageous for quick response teams to measure oil droplets underwater. Measuring oil droplet size distribution and concentration provides insight into the volume and spread of the oil. Real-time oil droplet analysis will aid in swift and timely decision-making. Processing holograms is an expensive operation in execution time and computer memory. Machine learning is one technique that can avoid the costly hologram reconstruction. Our work focused on training the models on real raw hologram images to test the use case of real-time data extraction. Our objective was to test the speed-up and precision trade-off achievable with different machine learning approaches. We tested two machine learning models, U-Net with a resnet34 backbone and YOLOv5 and then compared their performances. The U-Net approached the problem as a semantic segmentation problem. YOLOv5 approached the problem as a class localization problem. Particle statistics are of the number of oil particles and the size of each particle. We collected a dataset of real-world hologram images of oil particles. We trained our machine learning models on real-world raw data with labels derived from the hybrid method. Average precision and average recall were the performances used to compare models. All of the experiments and training were performed on a high-performance desktop equipped with GPU cards. We used average precision and average recall to evaluate the models. The findings suggested that the machine learning approaches were able to count the number of oil particles accurately, but were less precise when measuring the size of each individual particle. Compared to the hybrid method, we demonstrated a processing time reduction of 99%. YOLOv5 with GPU acceleration proved to be the superior model in all metrics. Machine learning is a promising technique for processing hologram images rapidly.