Image Reconstruction Performance Research Articles

DSU-Net: A Dynamic Stage Unfolding Network for high-noise image compressive sensing denoising

Deep unfolding networks (DUNs) have demonstrated considerable efficacy in the domain of compressive sensing (CS), attributed to their superior performance and interpretability. Nonetheless, many existing CS methodologies fail to account for the influence of noise prior to image sampling, resulting in blurring and distortion in the reconstructed images. To address these issues, this paper proposes a dynamic stage unfolding network (DSUNet). Firstly, a novel dynamic stage unfolding mechanism is proposed to achieve feature refinement by dynamically optimizing various noisy images in a sequential manner. Secondly, a step inertia fusion module (SIFM) is developed to execute multi-tiered information fusion from adjacent stages, thereby promoting feature reuse and minimizing the loss of detailed information. Finally, a step cross transformer denoiser (SCTD) is designed to capture correlations between distant pixels, effectively addressing the limitations associated with local attributes and significantly improving image reconstruction performance. Comprehensive experimental evaluations indicate that the proposed DSUNet achieves outstanding performance under both low CS ratios and high noise levels, successfully addressing the challenges of denoising and reconstruction in high-resolution and high-noise images. Codes are available at https://github.com/zzuli407/DSU-Net.

A Novel RPCA Method Using Log-Weighted Nuclear and L_(2,1) Norms Combined with Contrast-Limited Adaptive Histogram Equalization (CLAHE) for High Dimensional Natural and Medical Image Data

Estimating the true underlying images from distorted high-dimensional data is crucial for applications in high-profile fields such as crime detection in security, clinical settings and medical diagnosis in healthcare, and radar imaging in signal processing. Existing statistical methods often struggle with robustness and image reconstruction quality when processing high-dimensional image data. While Robust Principal Component Analysis (RPCA) is widely used for image recovery, its reliance on uniform weights with singular value decomposition (SVD) weakens performance, especially in noisy environments. The L1 norm also fails to capture image details and recovery under high noise levels, a critical limitation for applications like medical diagnoses, where detail is essential. These challenges emphasize the need for improved methods to handle noise and enhance image quality in sensitive fields. Therefore, this paper proposes a novel RPCA method that integrates CLAHE with Log weighted nuclear norm (LWNN) and the L2,1 norm for high-dimensional natural and medical imaging. To reduce the computational load, our novel method is formulated into a new optimization problem and solved using the Alternating Direction Method of Multipliers (ADMM). This method leverages LWNN for enhanced low-rank approximation to drastically prune out the anomalies in images and the norm for improved sparse component recovery. Our approach has superior performance in image reconstruction compared to other state-of-the-art methods (SOTAs), showing significant advancements with real-world datasets. An interesting finding of this research is that combining the LWNN with the L2,1 norm is highly effective at removing noise from images. Furthermore, when the CLAHE technique is combined with LWNN and the L2,1 norm, it significantly enhances the extraction of previously unseen features, making blood vessels in medical images much clearer and more distinguishable. This combination proves to be a powerful approach for medical image analysis, revealing details that are otherwise difficult to detect. This method will be used for crime detection in security intelligence, and clinical settings and medical diagnosis in human retinal eyes.

Color Reproduction of Chinese Painting Under Multi-Angle Light Source Based on BRDF

It is difficult to achieve high-precision color reproduction using traditional color reproduction methods when the angle is changed, and, for large-sized artefacts, it is also significantly difficult to collect a large amount of data and reproduce the colors. In this paper, we use three Bidirectional Reflectance Distribution Function (BRDF) modeling methods based on spectral imaging techniques, namely, the five-parameter model, the Cook–Torrance model and the segmented linear interpolation model. We investigated the color reproduction of color chips with matte surfaces and Chinese paintings with rough surfaces under unknown illumination angles. Experiments have shown that all three models can effectively perform image reconstruction under small illumination angle intervals. The segmented linear interpolation model exhibits a higher stability and accuracy in color reconstruction under small and large illumination angle intervals; it can not only reconstruct color chips and Chinese painting images under any illumination angle, but also achieve high-quality image color reconstruction standards in terms of objective data and intuitive perception. The best test model (segmented linear interpolation) performs well in reconstruction, reconstructing Chinese paintings at 65° and 125° with an illumination angle interval of 10°. The average RMSE of the selected reference color blocks is 0.0450 and 0.0589, the average CIEDE2000 color difference is 1.07 and 1.50, and the SSIM values are 0.9227 and 0.9736, respectively. This research can provide a theoretical basis and methodological support for accurate color reproduction as well as the large-sized scientific prediction of artifacts at any angle, and has potential applications in cultural relic protection, art reproduction, and other fields.

Double-image encryption and authentication scheme based on compressed sensing and double random phase encoding

To enhance the security of image information, a double-image encryption and authentication scheme combining compressive sensing (CS) and double random phase encoding (DRPE) is proposed. First, two plaintext images are taken as real and imaginary parts to form a complex-valued image, which is then encoded using DRPE. Next, extract the phase of the encrypted complex-valued image and encode it into authentication information. Simultaneously, the complex-valued image is sampled using the measurement matrix, which is optimized by Schmidt orthogonalization, and then quantized to form a compressed image. Subsequently, the authentication information is embedded into the compressed image, and by applying the permutation and diffusion algorithm to it, the ciphertext image is obtained. At the receiving end, the ciphertext is decrypted using inverse permutation and inverse diffusion algorithms, and the authentication information is decoded to acquire the authentication image. Finally, synchronous authentication of the two reconstructed images is achieved using a nonlinear cross-correlation method. Additionally, the keys in the proposed scheme are generated using high-dimensional chaotic systems, effectively reducing the required storage space and transmission bandwidth. Simulation results demonstrate that the proposed scheme has high image reconstruction performance at different compression ratios and possesses outstanding security and authentication capabilities.

Visually security privacy medical data protection scheme using compressive sensing and improved duffing chaotic system

Health and medical data frequently contain sensitive patient information that must be protected. Existing visual security schemes for medical images exhibit limitations in the imperceptibility of cipher images and the performance of image reconstruction. This paper introduces and evaluates a novel approach called Visual Meaningful Image Encryption (VMIE) for securing medical images. The proposed VMIE scheme employs a chaotic system based on the Duffing equation for initial encryption. Medical images are processed and encrypted in a sparse domain. A Bidirectional Chaotic Magic Transformation (BCMT) algorithm is then applied to scramble the sparse medical images. The scrambled data undergoes compression and diffusion. An adaptive embedding strategy employing the Discrete Cosine Stockwell Transform (DCST) integrates confidential data into the host image. The performance of the proposed chaotic system is validated through theoretical analysis and numerical simulation. Simulation results, along with comparisons to existing schemes, demonstrate the efficacy of the VMIE method in enhancing visual security and its suitability for natural images. The VMIE approach presented in this paper offers a promising solution for securing medical images, effectively addressing the limitations of current visual security schemes.

High-quality AFM image acquisition of living cells by modified residual encoder-decoder network

Atomic force microscope enables ultra-precision imaging of living cells. However, atomic force microscope imaging is a complex and time-consuming process. The obtained images of living cells usually have low resolution and are easily influenced by noise leading to unsatisfactory imaging quality, obstructing the research and analysis based on cell images. Herein, an adaptive attention image reconstruction network based on residual encoder-decoder was proposed, through the combination of deep learning technology and atomic force microscope imaging supporting high-quality cell image acquisition. Compared with other learning-based methods, the proposed network showed higher peak signal-to-noise ratio, higher structural similarity and better image reconstruction performances. In addition, the cell images reconstructed by each method were used for cell recognition, and the cell images reconstructed by the proposed network had the highest cell recognition rate. The proposed network has brought insights into the atomic force microscope-based imaging of living cells and cell image reconstruction, which is of great significance in biological and medical research.

Time-domain signal extraction and image reconstruction through scattering medium based on fully convolutional neural network

In scattering environments, laser-active imaging systems are affected by medium absorption and multiple scattering, which reduces the system's ability to effectively detect objects hidden behind the scattering medium. Conventional penetrating scattering-medium imaging methods are less effective under dynamic and strong-scattering conditions. Current deep learning-based image reconstruction techniques only perform image reconstruction from the perspective of the image, without taking the characteristics of the difference between the distribution of the signal and noise in the initially acquired time-domain signals into consideration. In this study, we start from the underlying time-domain distribution characteristics between the signal and the noise and use the full convolutional neural network to learn the time-domain distribution discrepancy between the scattering noise and the echo signal. We directly establish the mapping between noise-containing and noise-free signals to complete the extraction of the target reflective echo signals and to realize target image reconstruction in the scattering medium. We experimentally prove that the proposed method can realize the reconstruction of the target intensity and distance images in a strong scattering environment. Finally, we experimentally demonstrate that the convolutional neural network (CNN) maintains its image reconstruction capability for scattered images outside the range of the training data distribution.

Reconstructing the near- to mid-infrared environment in the stellar merger remnant V838 Monocerotis

Context. V838 Mon is a stellar merger remnant that erupted in a luminous red nova event in 2002. Although it has been well studied in the optical, near-infrared, and submillimeter regimes, its structure in the mid-infrared wavelengths remains elusive. Over the past two decades, only a handful of infrared interferometric studies have been performed, suggesting the presence of an elongated structure at multiple wavelengths. However, given the limited nature of these observations, the true morphology of the source has not yet been conclusively determined. Aims. By performing image reconstruction using observations taken at the VLTI and CHARA, we aim to map out the circumstellar environment in V838 Mon. Methods. We observed V838 Mon with the MATISSE (LMN bands) and GRAVITY (K band) instruments at the VLTI as well as the MIRCX/MYSTIC (HK bands) instruments at the CHARA array. We geometrically modelled the squared visibilities and the closure phases in each of the bands to obtain the constraints on the physical parameters. Furthermore, we constructed high-resolution images of V838 Mon in the HK bands using the MIRA and SQUEEZE algorithms to study the immediate surroundings of the star. Lastly, we also modelled the spectral features seen in the K and M bands at various temperatures. Results. The image reconstructions show a bipolar structure that surrounds the central star in the post-merger remnant. In the K band, the super-resolved images show an extended structure (uniform disk diameter ~1.94 mas) with a clumpy morphology that is aligned along a north-west position angle (PA) of −40°. On the other hand, in the H band, the extended structure (uniform disk diameter ~1.18 mas) lies roughly along the same PA. Yet the northern lobe is slightly misaligned with respect to the southern lobe, which results in the closure phase deviations. Conclusions. The VLTI and CHARA imaging results show that V838 Mon is surrounded by features resembling jets that are intrinsically asymmetric. This is further confirmed by the closure phase modelling. Further observations with VLTI can help to determine whether this structure shows any variations over time and also if such bi-polar structures are commonly formed in other stellar merger remnants.

Radial Undersampled MRI Reconstruction Using Deep Learning With Mutual Constraints Between Real and Imaginary Components of K-Space.

The deep learning method is an efficient solution for improving the quality of undersampled magnetic resonance (MR) image reconstruction while reducing lengthy data acquisition. Most deep learning methods neglect the mutual constraints between the real and imaginary components of complex-valued k-space data. In this paper, a new complex-valued convolutional neural network, namely, Dense-U-Dense Net (DUD-Net), is proposed to interpolate the undersampled k-space data and reconstruct MR images. The proposed network comprises dense layers, U-Net, and other dense layers in sequence. The dense layers are used to simulate the mutual constraints between real and imaginary components, and U-Net performs feature sparsity and interpolation estimation for k-space data. Two MRI datasets were used to evaluate the proposed method: brain magnitude-only MR images and knee complex-valued k-space data. Several operations were conducted for data preprocessing. First, the complex-valued MR images were synthesized by phase modulation on magnitude-only images. Second, a radial trajectory based on the golden angle was used for k-space undersampling, whereby a reversible normalization method was proposed to balance the distribution of positive and negative values in k-space data. The optimal performance of DUD-Net was demonstrated based on a quantitative evaluation of inter-method and intra-method comparisons. When compared with other methods, significant improvements were achieved, PSNRs were increased by 10.78 and 5.74dB, whereas RMSEs were decreased by 71.53% and 30.31% for magnitude and phase image, respectively. It is concluded that DUD-Net significantly improves the performance of MR image reconstruction.

Performance evaluation of deep learning image reconstruction algorithm for dual-energy spectral CT imaging: A phantom study.

To evaluate the performance of deep learning image reconstruction (DLIR) algorithm in dual-energy spectral CT (DEsCT) as a function of radiation dose and image energy level, in comparison with filtered-back-projection (FBP) and adaptive statistical iterative reconstruction-V (ASIR-V) algorithms. An ACR464 phantom was scanned with DEsCT at four dose levels (3.5 mGy, 5 mGy, 7.5 mGy, and 10 mGy). Virtual monochromatic images were reconstructed at five energy levels (40 keV, 50 keV, 68 keV, 74 keV, and 140 keV) using FBP, 50% and 100% ASIR-V, DLIR at low (DLIR-L), medium (DLIR-M), and high (DLIR-H) settings. The noise power spectrum (NPS), task-based transfer function (TTF) and detectability index (d') were computed and compared among reconstructions. NPS area and noise increased as keV decreased, with DLIR having slower increase than FBP and ASIR-V, and DLIR-H having the lowest values. DLIR had the best 40 keV/140 keV noise ratio at various energy levels, DLIR showed higher TTF (50%) than ASIR-V for all materials, especially for the soft tissue-like polystyrene insert, and DLIR-M and DLIR-H provided higher d' than DLIR-L, ASIR-V and FBP in all dose and energy levels. As keV increases, d' increased for acrylic insert, and d' of the 50 keV DLIR-M and DLIR-H images at 3.5 mGy (7.39 and 8.79, respectively) were higher than that (7.20) of the 50 keV ASIR-V50% images at 10 mGy. DLIR provides better noise containment for low keV images in DEsCT and higher TTF(50%) for the polystyrene insert over ASIR-V. DLIR-H has the lowest image noise and highest detectability in all dose and energy levels. DEsCT 50 keV images with DLIR-M and DLIR-H show potential for 65% dose reduction over ASIR-V50% withhigher d'.

Trusted Perception Method for Traffic Signs That Are Physically Attacked

Traffic sign recognition is a crucial method by which autonomous driving systems acquire road information, and is predominantly based on deep neural networks (DNNs). However, the recognition results of DNNs are not always trustworthy for traffic signs subject to abnormal disturbance. Recently, the phenomenon of adversarial examples successfully deceiving DNNs has garnered considerable attention. Because DNN-based computer vision techniques are becoming increasingly prevalent in traffic scenarios, the misclassification of attacked traffic signs by DNN classifiers poses serious safety hazards. Although numerous methods have been proposed for crafting physical adversarial examples that are robust in the real world, most existing defense approaches focus on digital attacks, which necessitate the adversary infiltrating the embedded system; thus, it becomes challenging to obtain results. A reliable approach for defending against physical adversarial traffic signs enables autonomous vehicles to achieve trusted perception of traffic signs. In this paper, we present a deep image prior-based pipeline to defend against robust adversarial traffic signs in the real world, an approach that circumvents the need for prior data sets during training. Our approach protects the safety of autonomous vehicles by performing image reconstruction of captured traffic sign images. The genuine traffic sign class can be inferred by leveraging the consistency of the victim classifier’s decision results for reconstructed images at different stages. Additionally, we evaluate the efficacy of our defense pipeline for detecting other potential types of physical adversarial traffic signs that may exist in the real world, thus demonstrating the generalizability of our approach.

Electrode Placement Optimization for Electrical Impedance Tomography Using Active Learning

Electrical impedance tomography (EIT) offers a versatile imaging modality with a multitude of applications, although it encounters accuracy limitations. Herein, a novel systematic framework is presented that integrates a neural network (NN), active learning, and transfer learning to optimize electrode placement, improving image reconstruction performance based on user‐defined metrics. Given the many‐to‐one mapping between electrode configuration and the performance metric, the approach utilizes a NN that predicts performance metrics from electrode placement input. To maintain NN's prediction accuracy for unseen electrode configurations, performance metrics are maximized while iteratively updating the NN via active learning during the optimization process. Transfer learning is employed to expedite optimization of electrode placements for time‐consuming iterative reconstruction techniques by fine‐tuning a NN initially trained on one‐step reconstruction data. The method is validated using two representative reconstruction methods: one‐step reconstruction with Newton's one‐step error reconstructor prior and the iterative total variation method. This research underscores the potential of the proposed framework in addressing EIT's inherent limitations and augmenting its performance across diverse applications and reconstruction methods. The framework could potentially contribute to the advancement of noninvasive medical imaging, structural health monitoring, strain sensing, robotics, and other fields that depend on EIT.

Image Reconstruction Requirements for Short-Range Inductive Sensors Used in Single-Coil MIT.

MIT (magnetic induction tomography) image reconstruction from data acquired with a single, small inductive sensor has unique requirements not found in other imaging modalities. During the course of scanning over a target, measured inductive loss decreases rapidly with distance from the target boundary. Since inductive loss exists even at infinite separation due to losses internal to the sensor, all other measurements made in the vicinity of the target require subtraction of the infinite-separation loss. This is accomplished naturally by treating infinite-separation loss as an unknown. Furthermore, since contributions to inductive loss decline with greater depth into a conductive target, regularization penalties must be decreased with depth. A pair of squared L2 penalty norms are combined to form a 2-term Sobolev norm, including a zero-order penalty that penalizes solution departures from a default solution and a first-order penalty that promotes smoothness. While constraining the solution to be non-negative and bounded from above, the algorithm is used to perform image reconstruction on scan data obtained over a 4.3 cm thick phantom consisting of bone-like features embedded in agarose gel, with the latter having a nominal conductivity of 1.4 S/m.

Quaternion fast and accurate polar harmonic Fourier moments for color image analysis and object recognition

Image moments, as a kind of global feature descriptor of images, have become a valuable tool for pattern recognition and image analysis. However, traditional methods are mainly used to deal with grayscale images. In this paper, we apply quaternions to fast and accurate polar harmonic Fourier moments, proposing a kind of quaternion fast and accurate polar harmonic Fourier moment (QFAPHFM) capable of handling color images. Furthermore, this paper provides a detailed analysis of the invariance of QFAPHFMs under rotation, scaling, and translation transformations. The experimental results show that QFAPHFMs exhibit excellent performance in both image reconstruction and object recognition tasks. QFAPHFMs achieve accurate image reconstruction under noiseless and noisy conditions, and demonstrate excellent recognition performance in the color-based object recognition tasks.

Rotational Convolution Design in Convolutional Neural Networks for Direct 3D Electromagnetic Tomography

The three-dimensional (3D) reconstruction of Electromagnetic Tomography (EMT) is an important task for many applications, such as the non-destructive testing of inner defects in rail systems. Additionally, image reconstruction algorithms utilizing deep learning methods have been verified to be useful in recent years. Therefore, the interpretability of deep learning is a question that is relevant to its application in other areas. This paper proposes an innovative rotational convolution pattern, Conv-P, for convolutional neural network (CNN) image reconstruction in a 3D EMT system. This pattern is based on the projection relationships inherent in tomographic imaging, where each convolution is performed on adjacent projections along the excitation rotation direction. The advantage of this pattern is that it can generate the convolution process by utilizing the 3D structural information from real sensors. To verify the effectiveness of this convolution pattern, we constructed a 3D dual-layer 16-coil EMT model and tested its image reconstruction performance. The results demonstrate that, compared with two common convolution patterns, Conv-P achieves a 4.7% and 4.1% increase in the Image Correlation Coefficient (CC), a 19.8% and 13.1% reduction in the Relative Image Error (IE), a 0.67% and 1.59% increase in the Peak Signal-to-Noise Ratio (PSNR), and a 3.24% and 0.74% increase in the Structural Similarity Index Measure (SSIM).

DLGAN: Undersampled MRI reconstruction using Deep Learning based Generative Adversarial Network

Magnetic Resonance Imaging (MRI) is a crucial tool for quantitative image analysis and clinical diagnosis, providing detailed anatomical images to assist in the detection of various abnormalities. However, the widespread use of MRI is hindered by the challenges associated with long sampling periods and low-resolution image processing, prompting the development of various traditional methods to address these limitations. In recent times, advanced Deep Learning (DL) techniques have been applied to tackle the inverse problem of reconstructing MRI images from undersampled data. These DL models have demonstrated substantial improvements in terms of image reconstruction performance, cost-effectiveness, and reduced acquisition time, offering significant potential for further enhancements in this domain. This study introduces a novel DLGAN (Deep Learning Generative Adversarial Network) model comprising two sub-GAN modules tailored for specific datasets. Each module incorporates DLGAN (Generator Block, GB) and DLGAN (Discriminator Block, DB) blocks, strategically designed to regenerate MR images from K Space data. These blocks effectively leverage information from both the ground truth and KSpace characteristics, resulting in enhanced image reconstruction performance while also reducing complexity and improving overall efficiency. The DLGAN model effectively addresses common issues such as artefact removal and vanishing-gradient problems through the extraction of hierarchical features. To evaluate the model’s effectiveness, comprehensive experiments were conducted, utilizing metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The experimental findings demonstrate that the proposed DLGAN model outperforms the most recent designs in MRI image reconstruction, showcasing its potential for significant advancements in clinical diagnosis and quantitative image analysis.

ISP-IRLNet: Joint optimization of interpretable sampler and implicit regularization learning network for accerlerated MRI

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) was proposed to accelerate data acquisition and reconstruct MR images from under-sampled data in k-space. However, the traditional approaches design the sampling patterns separately from the reconstruction process, which often leads to suboptimal reconstruction performance. To address this issue, we propose a joint optimization model dubbed as ISP-IRLNet to yield optimal recovery including an interpretable sampler and a reconstructor using the implicit regularization learning network. In particular, we introduce a probabilistic distribution model of the sampling incorporated into the prior of k-space locations, which provides a robust CS-MRI sample pattern with an explicit expression. To estimate the approximating gradient in backward propagation for more stable training, we employ a differentiable binarization strategy. In addition, we unroll the ADMM optimization for solving the regularized reconstruction model to be a deep network (i.e., IRLNet) with a learnable implicit regularization sub-network to learn a regularizer or prior term implicitly. The experiments on our collected brain and public knee datasets demonstrate that our method provides an optimal sample pattern and achieves superior image reconstruction performance compared with existing methods, especially at high acceleration rates.

A parametrization study of TEULM’s image reconstruction performance

Time efficient ultrasound localization microscopy (TEULM) allows imaging the microvascular structures with reduced frame rates compared to standard ULM. The quality of reconstructed images relies on different parameters. In this study, we aim at understanding the links between TEULM performance and three key parameters: SVD k-value (representing the number of components considered as noise), the maximum linking distance (indicating the assumed maximum distance that a microbubble can cover between two successive frames), and the minimum length of a track. We generated TEULM density and dynamics maps, singularly varying these three parameters. Next, we evaluated the quality of TEULM images by means of a root mean squared error (RMSE), dice score, and Fourier ring correlation (FRC) analysis. A publicly available preclinical in-vivo dataset was used for this study. Results demonstrate that minimum track length and k-value have the greatest impact on both density and dynamic maps RMSE, whereas the minimum linking distance is the key factor influencing the dice score. Additionally, the relationship between these parameters and metrics is non-linear. Furthermore, the resolution measured by the FRC is negatively affected by increasing the SVD k-value and by increasing the maximum linking distance but positively affected by increasing the minimum track length.

NesTD-Net: Deep NESTA-Inspired Unfolding Network With Dual-Path Deblocking Structure for Image Compressive Sensing.

Deep compressive sensing (CS) has become a prevalent technique for image acquisition and reconstruction. However, existing deep learning (DL)-based CS methods often encounter challenges such as block artifacts and information loss during iterative reconstruction, particularly at low sampling rates, resulting in a reduction of reconstructed details. To address these issues, we propose NesTD-Net, an unfolding-based architecture inspired by the NESTA algorithm, designed for image CS. NesTD-Net integrates DL modules into NESTA iterations, forming a deep network that continuously iterates to minimize the l1 -norm CS problem, ensuring high-quality image CS. Utilizing a learned sampling matrix for measurements and an initialization module for initial estimate, NesTD-Net then introduces Iteration Sub-Modules derived from the NESTA algorithm (i.e., Yk , Zk , and Xk ) during reconstruction stages to iteratively solve the l1 -norm CS reconstruction. Additionally, NesTD-Net incorporates a Dual-Path Deblocking Structure (DPDS) to facilitate feature information flow and mitigate block artifacts, enhancing image detail reconstruction. Furthermore, DPDS exhibits remarkable versatility and demonstrates seamless integration with other unfolding-based methods, offering the potential to enhance their performance in image reconstruction. Experimental results demonstrate that our proposed NesTD-Net achieves better performance compared to other state-of-the-art methods in terms of image quality metrics such as SSIM and PSNR, as well as visual perception on several public benchmark datasets.

Reconstruction of degraded image transmitting through ocean turbulence via deep learning

When a laser carrying image information is transmitted in seawater, the presence of ocean turbulence leads to significant degradation of the received information due to the effect of interference. To address this issue, we propose a deep-learning-based method to retrieve the original information from a degraded pattern. To simulate the propagation of laser beams in ocean turbulence, a model of an ocean turbulence phase screen based on the power spectrum inversion method is used. The degraded images with different turbulence conditions are produced based on the model. A Pix2Pix network architecture is built to acquire the original image information. The results indicate that the network can realize high-fidelity image recovery under various turbulence conditions based on the degraded patterns. However, as turbulence strength and transmission distance increase, the reconstruction accuracy of the Pix2Pix network decreases. To further improve the image reconstruction ability of neural network architectures, we established three networks (U-Net, Pix2Pix, and Deep-Pix2Pix) and compared their performance in retrieving the degraded patterns. Overall, the Pix2Pix network showed the best performance for image reconstruction.