How convolutional-neural-network detects optical vortex scattering fields

Abstract

Light scattering through disordered media is a critical topic in optical engineering as its ubiquity in natural and artificial systems. Recent progress has shown that deep learning is capable to recognize topological charge values carried by orbital angular momentum (OAM) waves with ultra-fine resolution under scattering environment. However, the physical mechanism of how a deep learning convolutional neural network (CNN) fulfills such tasks remains unclear. In this paper, in perspective of optical vortex scattering field detection, we studied the basic physical mechanism of the CNN on recognizing scattered vortex beams carrying OAMs. It has been demonstrated that a CNN uses statistical invariance of both spatial phase front of an incident OAM wave and intrinsic features of a specific disordered medium across large-scale datasets to identify the OAM topological charge values from speckles. This work can provide insightful reference for CNN-assisted OAM-based scattering detection.