Deep convolutional neural network for binary regression of three-dimensional objects using information retrieved from digital Fresnel holograms

• Published in: Applied physics. B, Lasers and optics Lasers and optics, 2022-08, Vol.128 (8), Article 157
• Main Authors: Mahesh, R. N. Uma, Nelleri, Anith
• Format: Article
• Language: English
• Subjects: Applied physics; Artificial neural networks; Datasets; Decision trees; Digital imaging; Engineering; Errors; Histograms; Holograms; Information retrieval; Lasers; Machine learning; Multilayer perceptrons; Multilayers; Neural networks; Object recognition; Optical Devices; Optics; Photonics; Physical Chemistry; Physics; Physics and Astronomy; Quantum Optics; Regression; Support vector machines
• ISSN: 0946-2171; 1432-0649
• DOI: 10.1007/s00340-022-07877-w

A deep convolutional neural network (CNN)-based binary regression task on 3D objects using concatenated intensity-phase (whole information) image dataset retrieved from experimentally generated off-axis digital Fresnel holograms is utilized in this work. Images in the dataset were prepared using the intensity and the phase (depth) information retrieved computationally from the hologram presented as a single image in a concatenated fashion, which accommodates the whole information of the 3D object. The data set comprises 2268 images of the chosen eighteen objects at different recording distances and various rotation angles. Binary regression task in deep learning done on holographic information of 3D objects is equivalent to the 3D objects prediction done on whole information objects data set, which produces continuous labels as output, justifies the intention of the present work. The loss, mean square error (MSE), and mean absolute error (MAE) calculations show less error on the training set. CNN has a good performance on the test set and a constant performance on the validation set with R 2 score (coefficient of determination) values of 0.73 and 0.25 and explained variance (EV) regression score values of 0.83 and 0.30. Further, CNN has the higher values of R 2 score and EV regression score as compared to the K-nearest neighbor (KNN), support vector machine (SVM), multi-layer perceptron (MLP), decision tree (DT), AdaBoost (ADB), random forest (RF), extra trees (ET), gradient boosting (GB), histogram gradient boosting (HGB), and stochastic gradient descent (SGD) regressors on the test/validation sets. The results from the proof of concept experiment suggest the efficacy of the proposed method.