In this study, we propose a method for estimating the 3D position and diameter of individual particles from single-view 2D particle distribution images using deep learning. To train the model, we generated a synthetic dataset of particle images using Blender, a 3DCG software, under realistic optical conditions including focal length, sensor size, and depth of field. Particle diameters ranged from 20 μm to 100 μm, and particles were placed at various positions within a defined 3D space. Each image contained a single particle, which was cropped and resized to 32×32 pixels as input to a ResNet-based regression model. The model was trained to predict the particle's depth and diameter. Evaluation using test data showed that the model achieved high accuracy, with mean absolute errors of 0.0159 for depth and 0.0147 for diameter (normalized values). These results demonstrate that even from a single 2D image, the particle's spatial geometry can be estimated with high precision by leveraging visual features such as blur, brightness, and apparent size.
TODA et al. (Wed,) studied this question.