Deep convolutional neural network for three-dimensional objects classification using off-axis digital Fresnel holography

Abstract

A Deep Convolutional Neural Network (CNN)-based three-dimensional (3D) objects classification using an augmented holographic phase image dataset is proposed. Off-axis Fresnel digital holography is used for retrieving 3D object information as phase images. The phase image dataset of the five different 3D objects considered in this study is prepared under various recording distances and rotational angles. The CNN based architecture classifies the 3D objects into two sets comprising of TRUE and FALSE classes with higher accuracy and minimal loss on the training set for the phase image dataset. The CNN has achieved higher precision, lower recall, and higher F1-score for the TRUE class on the test set. The experimental results demonstrate that the proposed CNN is capable of classifying 3D objects with an overall accuracy of 86% and AUC of 96% for the phase image dataset. The feasibility and classification performance are deduced from a proof of the concept experiment.