Unsupervised deep learning for single-shot incoherent holographic 3D imaging

Paper Information

Background Introduction

Digital holography has attracted much attention for its ability to image 3D scenes from a single viewpoint. Compared with direct imaging, digital holography is an indirect multi-step imaging process that includes optical recording of holograms and numerical computational reconstruction, providing a wide range of application scenarios for computational imaging methods including deep learning. In recent years, incoherent digital holography has attracted much attention due to its high imaging resolution, absence of scattering noise and edge effects, and low cost. Currently, incoherent holography has been applied to aperture imaging, super-resolution imaging, large depth-of-field imaging and lattice light sheet microimaging.

In recent years, deep learning has been applied to incoherent digital holography. However, all current reports are based on data-driven supervised learning methods, which require large amounts of pairwise labeled data and suffer from problems such as insufficient generalization. To address the above challenges, this paper proposes a one-shot non-coherent holographic self-calibrating 3D reconstruction method without training neural network a priori, called SC-RUN. SC-RUN improves the fidelity and signal-to-noise ratio of the point-spread function (PSF), and achieves a high-fidelity and artifact-free reconstruction of 3D objects with only a single hologram. In this paper, the effect of SC-RUN is clearly demonstrated using interference-free coded aperture correlation holography (I-COACH) imaging as an example.

Methodological principles

Fig. 1 Non-interference coded aperture correlation holography device

Light from an incoherent light source is focused by lens L1 to illuminate an object. The object is located near the front focal plane Z3 of the lens L2, allowing the object to be considered to be located in the far field of the CPM. The SLM loaded with the encoded phase is located at the distance d from the lens L2, and the SLM is preceded by a polarizer P. Since the imaging model of I-COACH is linearly spatially invariant in intensity, the hologram of the object recorded by the sensor can be regarded as the incoherent intensity superposition of the holograms of innumerable object points, and thus the light field of an object point can be theoretically analyzed and then convolved or superimposed to obtain the multiobject object imaging model.

Fig. 2 Structure of SC-RUN-calibration point diffusion function

Fig. 3 Structure of SC-RUN-single-shot imaging based on untrained neural network prior

Fig. 4 SC-RUN-single-shot 3D imaging structure based on untrained neural network prior

System Optical Path

The multi-channel I-COACH experimental system is shown in Fig. 5, in which the product parameters of the amplitude-type spatial light modulator are shown in the following table.

Figure 5 I-COACH experimental setup

The spatial light modulator used in this experiment is our TSLM07U-A, and its parameters and specifications are as follows:

The system consists of two target channels in different axial planes, where a digital micromirror device (DMD) is used as target 1 in channel 1 and an amplitude-type spatial light modulator is used as target 2 in channel 2. Light from a spatially incoherent light-emitting diode (LED) is collected by a concentrator to illuminate the object, and then the diffracted object light in the two channels is combined by a beam-splitting prism (BS1) and collimated through a lens L for collimation. The polarizer P polarizes the object light in the direction of the modulation axis of the pure phase SLM. Finally, the light wave modulated by the pure phase SLM is recorded by a CMOS sensor. The pure-phase SLM is loaded with a hologram synthesized by the GSA algorithm.

Results

Fig. 6 Calibration results of SC-RUN for PSF. a) Hologram, b) Original PSF, c) Results of nonlinear reconstruction using the original PSF, d) Known object, e) Calibrated PSF, and f) Results of nonlinear reconstruction using the calibrated PSF.

Fig. 7 2D experimental results of SC-RUN and nonlinear reconstruction

Fig. 8 3D experimental results of SC-RUN and nonlinear reconstruction

The above experimental results show that SC-RUN performs well on I-COACH, thus indicating that this strategy of pre-calibrating the PSF and then reconstructing the object via neural networks has great potential. Currently, many optical imaging techniques are implemented by designing specialized PSFs. For example, sub-diffraction limit point PSFs are generated by wavefront coding to enable super-resolution imaging. Similarly, imaging depth can be extended by using wavefront coding to make the PSF insensitive to mis-focusing. For other information, such as the depth, spectrum and polarization of the object, it can be encoded into the PSF to increase the imaging dimension. The above computational imaging techniques rely heavily on the a priori information of the PSF, and SC-RUN allows obtaining PSFs with high fidelity and high signal-to-noise ratios.Therefore, excellent reconstruction results can be obtained when the forward operator is known. Furthermore, since SC-RUN enforces measurement consistency without the need for datasets and labels, and given that most imaging tasks involve one or more inverse solution models with known forward operators, SC-RUN can be easily applied to a variety of other imaging tasks.

Paper Summary

In this paper, we proposed a generalized unsupervised incoherent holographic 3D reconstruction framework, SC-RUN, which combines the physical knowledge of nonlinear reconstruction methods and forward imaging models to perform the reconstruction task via a neural network with additional physical constraints.SC-RUN takes both temporal resolution and fidelity into account, is robust, and does not require much labeled data-driven information. In addition, experimental results show that high-fidelity reconstruction of complex objects with intensity variations is achieved for the first time in incoherent holography.SC-RUN is generally suitable for a variety of optical configurations and is easily adaptable to other imaging tasks. In addition, SC-RUN has a wide range of potentials for super-resolution imaging, aperture imaging, depth-of-field extended imaging, and multidimensional information multiplexing, paving the way for obtaining multidimensional information in dynamic optical fields.