Particle size and spatial distribution measurement using dual-view lens and generative adversarial networks

Visual measurement offers the intuitive advantage of “what you see is what you get” in particle characterization. As a technique featuring a simple experimental setup, shadow imaging captures particle silhouettes and extracts particle diameter and spatial position from defocus features such as blur extent and contour sharpness. However, conventional approaches often suffer from non-unique mappings between defocus features and physical parameters, which undermines measurement accuracy and reliability. Moreover, the limited field of view under high-resolution imaging restricts statistical representativeness when characterizing particle groups. To overcome these limitations, we encode defocus patterns using Fourier series descriptors and employ a generative adversarial network (GAN) to learn the mapping between particle images captured at low (1.92×) and high (8.86×) magnifications. Training was conducted on 500 paired images using an NVIDIA GeForce RTX 4090 GPU (24 GB VRAM). Once trained, the model enables high-accuracy inference of particle diameter and depth position from low-resolution inputs. Experiments using polystyrene particles ranging from 30–120 µm demonstrate that our method achieves mean absolute percentage errors of 5.59% for diameter and 0.19 mm for spatial position. Compared to direct low-resolution measurements, our framework reduces the mean relative error and standard deviation of absolute error to 32.7% and 64.1%, respectively. This approach supports large-field, high-throughput particle imaging and shows potential for real-time concentration field mapping and particle dispersion analysis.