Recognition of Multiscale Dense Gel Filament-Droplet Field in Digital Holography With Mo-U-Net

Accurate particle detection is a common challenge in field characterization with digital holography, especially for gel secondary breakup dense complex particles and filaments of multi-scale strong background noises. This study proposes deep learning method called Mo-U-net which adapted from the combination U-net Mobilenetv2, demostrates its application to segment filament-droplet drop. Specially, pruning applied on Mo-U-net, cuts off about two-thirds layers save training time while remaining high segmentation accuracy. The performances are quantitatively evaluated by three indices, positive intersection over union (PIOU), average square symmetric boundary distance (ASBD) diameter-based prediction statistics (DBPS). experimental results show that area accuracy (PIOU) reaches 83.3 % , 5 higher than adaptive-threshold (ATM). error only one pixel-wise length, third ATM. And also shares coherent size distribution (DBPS) droplet diameters reality. These demonstrate recognition capability providing accurate statistical data variety holographic diagnostics. Public model address: https://github.com/Wu-Tong-Hearted/Recognition-of-multiscale-dense-gel-filament-droplet-field-in-digital-holography-with-Mo-U-net .