Convolutional neural network for 2D-reconstructions of rough particles from their experimental speckle images in interferometric particle imaging

Abstract

A convolutional neural network (CNN) has been trained to reconstruct the two-dimensional (2D) shape of rough particles from their interferometric defocused images. The experimental set-up uses a digital micro-mirror device (DMD). The particles programmed on the DMD can have different morphologies: stick, cross, dendrite, L, T and Y. Sticks, crosses and dendrites are centrosymmetric, while l-, T-, and Y-shaped particles are non-centrosymmetric. For each family of particle, the training of the CNN has been performed for particle's sizes that cover a decade from hundreds of micrometers to millimeters, and for an arbitrary 3D-orientation of the particle. Accurate reconstructions and particle sizing can be done. Using three orthogonal views of the same particle, a three-dimensional (3D) reconstruction can be further done to estimate the 3D-morphology of the particle.