|A machine learning algorithm for accurately reconstructing 3D particle fields.|
- Particle field imaging using holography
- Size distribution of droplets in fuel and agricultural sprays
- Imaging through diffused media, light field imaging, defocus imaging
- Fluid dynamics, crystallization monitoring, environmental science
- Biophysical and medical studies such as aerosol flow, microfluidics, brain activities
Key Benefits & Differentiators
- Wide range of particle concentrations: concentrations up to 0.061 ppm - 305 times previous demonstrations - has been tested
- High extraction rate: up to 94%
- High positioning accuracy (error of
- High speed: 30x faster than the analytical RIHVR method
- Not built based on models: generalizable algorithm and reduced training requirements for new hologram datasets
Researchers in Prof. Jiarong Hong’s laboratory have developed an image reconstruction algorithm using a machine learning approach for accurate reconstruction of three-dimensional particle fields from digital holography. Image reconstruction algorithms are used to extract useful particle information (such as size and 3D position of bubbles, aerosols, cells, etc.) encoded as complex interference patterns in digital holograms. However, currently available algorithms perform suboptimally at realistic conditions such as when high particle concentrations, high dynamic, background or cross-interference noises are present. Moreover, practically relevant particle field reconstruction often requires sophisticated data acquisition, tedious fine tuning and is computationally intensive.
Using specialized U-net architecture, the algorithm disclosed here has been shown to accurately reconstruct images and extract particle information with high prediction accuracy and extraction rate at significantly wider range of concentrations than previously demonstrated. Particularly, instability issues and reduced localization accuracy caused due to sparsity in the particle field is tackled in this novel algorithm. The algorithm is developed using a combination of synthetic and experimental data, and is optimized for quickly producing high localization accuracy, smooth background and reducing ghost particles. The design of this system reduces the need to fully learn the required physics, therefore reduces the training and tuning requirements for new hologram datasets and is easily adaptable in a wide range of applications. In other words, this learning-based algorithm is highly generalizable. Lastly, this learning-based hologram reconstruction is >30 times faster than currently available methods, making it suitable for developing systems with real-time reconstruction capabilities.
Phase of Development
Pilot scale demonstration in processing high density holograms.
Ready for Licensing
This technology is now available for license! The University is excited to partner with industry to see this innovation reach its potential. Please contact Doug Franz to share your business’ needs and your licensing interests in this technology. The license is for the sale, manufacture or use of products claimed by the patents.
- Machine learning holography for 3D particle field imaging. Optics Express 28.3 (2020): 2987-2999.
- Machine learning holography for measuring 3D particle size distribution arXiv preprint arXiv:1912.13036 (2019).