Autofocusing algorithm for a digital holographic imaging system using convolutional neural networks

Abstract

Digital holographic imaging systems are promising three-dimensional imaging systems that acquire holograms via interference of a reference wave and an object wave. Using digital holography and the numerical diffraction theory, an image can be reconstructed at any distance from the hologram. However, accurate determination of the distance of the object from the hologram is required to focus the image. Various autofocusing algorithms have been studied. The conventional autofocusing algorithm creates the focused image by evaluating iteratively reconstructed images using focus metrics. Owing to the iterative image reconstruction process, the computational time is very long. In this paper, an autofocusing algorithm for a digital holographic imaging system using convolutional neural networks, similar to pattern recognition systems, is proposed. Using the proposed method, the distance of the object from the hologram is obtained more rapidly than using the conventional method.

On the use of deep learning for phase recovery

“…Furthermore, they revealed the advantages of neural networks over other machine learning algorithms in the task of autofocusing. Immediately afterward, Son et al 102 also verified the feasibility of autofocus by classification through numerical simulations. Subsequently, Couturier et al 103 improved the accuracy of defocus distance estimation by using a deeper CNN for categorizing defocus distance into a greater number of classes.…”

“…: 485,856 pairs Cross entropy Ren et al 101 Hologram Defocus distance (5 types) CNN Expt. : >5000 pairs Cross entropy Son et al 102 Hologram Defocus distance (10 types) CNN Sim. : 40,180 pairs Cross entropy Couturier et al 103 Hologram Defocus distance (101 types) DenseNet Expt.…”

Robust and fast auto-focusing using convolutional neural networks for off-axis digital holography

“…14) Thus, various methods using CNNs for autofocusing in DH have recently been proposed. [15][16][17][18] Pitkaaho et al chose a CNN approach in which the architecture is based on AlexNet, which won the 2012 ImageNet Large-Scale Visual Recognition Challenge (ILSVRC). Since this architecture is designed to classify discontinuous label values in training and validation, it can only estimate the trained label output instead of the continuous regression value.…”

“…They trained their network with a large number of experimental images, which were obtained in phase-shift optical configuration, and achieved better distance estimation accuracy in real-time. 18) Despite insignificant improvements in distance estimation through deep learning, the studies mentioned above have not been evaluated for their robustness under untrained conditions. To apply it to actual optical applications, it should be able to achieve sufficient performance within the tolerance required by the product.…”