Session Index

This paper proposes a conceptual design for hybrid optical-electronic convolution neural network. The key design exploits the inherent advantage of 3D photonics architecture that provides inversion, multiplication, and summation of the images. This concept can be made possible by using today smart phones or tablets and high-definition liquid crystal displays.

 

Computation resources and machine learning have made a rapid progress in the last decade. Inspired by the computer science community focusing on data interpretation, computational optics took further step to investigate more possibility about the image formation. In this study, I will present some data-driven works including image denoising, phase retrieval and unconventional imaging system based on deep learning neural network.

 

A prism-hologram-prism sandwiched recording method is proposed for the fabrication of polarization-selective substrate-mode volume holograms with a large diffraction angle. The method belongs to a technique of longer-wavelength construction for shorterwavelength reconstruction and is much easier than that of traditional method which has high application potential in holographic photonics.

 

A novel deep learning model for particle detection and sorting in three-dimensions with digital holography was proposed and demonstrated. Deep learning models such as U-net and convolution neural network are implemented to predict and estimate the particle’s size and its orientations in lateral and axial directions.

 

In this study, the presented novel algorithm, based on the interferometry techniques, is used to rebuild the profile of the step height sample and the dirty position. By contrast, the phase unwrapping algorithm and the zero-order interference fringe fail to simultaneously rebuild the profile of sample and the dirty position.

 

This study presents the algorithm to distinguish the strong and signals, and the noise in the input of three images. All of the detected strong signal (or weak signal) can combine to the one stitching image. By contrast, the known Speeded Up Robust Features does not effectively distinguish these signals.