Enhancement Method of Image Quality for Lensless Imaging Based on Physics-Driven Unsupervised Learning
Objective Digital holography is an optical imaging technique that records and reconstructs the complete wavefront information of an object field.In-line holography,known for its broader spatial bandwidth and simpler imaging system,has gained widespread attention and application.However,traditional in-line holography suffers from interference by zero-order and twin images during object field reconstruction,affecting the observation and measurement of the object field's complex amplitude.In addition,traditional phase recovery algorithms can produce significant errors,especially under high-resolution and large field-of-view imaging requirements,where reconstruction quality can be compromised.Inaccurate initial estimates or incomplete prior information may cause the algorithm to fail to converge or produce incorrect results due to the strong reliance of phase retrieval methods on initial estimates and prior information.Data-driven neural network methods are limited by dataset collection and quality and lack interpretability.Physics-driven neural networks combine neural networks with models that adhere to physical laws,mitigating the drawbacks of data-driven neural networks.However,physically motivated neural networks require certain constraints on data acquisition when only a single hologram is provided as input.Consequently,the final reconstruction results are susceptible to the influence of secondary interference fringes,leading to uneven backgrounds in the recovered results.In response to these issues,a method for reconstructing holographic images using a dual-input physics-driven neural network with amplitude constraint(DPNNA)is proposed.Methods Firstly,the DPNNA method is constructed.Two holograms captured at diffraction distances d1 and d2 are used as inputs to the neural network.A physical model is employed to compute the estimated hologram and phase corresponding to diffraction distance d1 based on the neural network's estimated phase.Subsequently,combining the estimated phase at d1 with the true amplitude,the physical model generates the estimated hologram corresponding to diffraction distance d2.The loss function is computed using the estimated and true holograms,and network parameters are optimized to achieve phase or amplitude imaging.Then,in-line holograms of phase and amplitude resolution targets,as well as phase objects with irregular phase variations,are simulated to validate the feasibility of DPNNA method.The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)among the reconstruction results of traditional Gerchberg-Saxton(G-S)method,the single-input physics-driven neural network method PhysenNet,and the proposed DPNNA method,and the ground truth are calculated.Finally,a lensless imaging optical setup is constructed to record in-line holograms containing information on different objects.The DPNNA method is employed for reconstruction and compared with other in-line holographic reconstruction methods.Results and Discussions The DPNNA method is utilized to simulate the reconstruction of pure phase objects with different phase distributions and is then compared with the reconstruction results obtained using the G-S algorithm and PhysenNet(Figs.5-7).Overall,when compared to the traditional G-S method and PhysenNet,the proposed DPNNA method demonstrated significant advantages in the high-quality recovery of object amplitude and phase information.This is particularly notable in lensless imaging,where it achieves the highest evaluation metrics.Subsequently,numerical reconstruction of holographic images is performed using in-line holograms of phase and amplitude resolution targets recorded in experiments(Figs.8 and 10).The results show that using the DPNNA method for holographic image reconstruction far outperforms other algorithms,with the weakest background noise in the reconstructed images.The DPNNA method not only effectively removes twin images but also demonstrates the robustness for different diffraction distances and noise levels(Tables 3 and 4).Conclusions We propose a DPNNA method.By combining the dual-input physics-driven neural network with the G-S reconstruction algorithm,accurate phase or amplitude reconstruction can be achieved without requiring a large amount of training data,ensuring stability and accuracy.Compared to the traditional G-S algorithm,the DPNNA method performs better in handling edge information and exhibits weaker background noise.The DPNNA method demonstrates good reconstruction accuracy for both amplitude-type and various phase-type objects,with stronger generalization ability and interpretability compared to data-driven neural network methods.The proposed method provides a low-cost,high-precision solution for phase imaging,overcoming dependencies on initial estimates and prior information,as well as issues such as high-cost dataset construction and uneven backgrounds.Combining a dual-input physics-driven neural network with amplitude constraint leads to more accurate in-line holographic imaging.This advancement holds significant application value in the field of computational imaging.
NETL
NSTL
万方数据
