Model-Informed Multistage Unsupervised Network for Hyperspectral Image Super-Resolution

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2024, Volume and Issue: 62, P. 1 - 17

Published: Jan. 1, 2024

By fusing a low-resolution hyperspectral image (LrMSI) with an auxiliary high-resolution multispectral (HrMSI), super-resolution (HISR) can generate (HrHSI) economically. Despite the promising performance achieved by deep learning (DL), there are still two challenges remaining to be solved. First, most DL-based methods heavily rely on large-scale training triplets, which reduces them limited generalization and poor practicability in real-world scenarios. Second, existing pursue higher designing complex structures from off-the-shelf components while ignoring inherent information degradation model, hence leading insufficient integration of domain knowledge lower interpretability. To address those drawbacks, we propose model-informed multi-stage unsupervised network, M2U-Net for short, leveraging both prior (DIP) model information. Generally, is built three-stage scheme, i.e., (DIL), initialized establishment (IIE), generation (DIG) stages. The first stage exploit via tiny network whose parameters outputs will serve as guidance following Instead feeding uninformed noise input three, IIE aims establish expressive HrHSI-relevant resorting spectral mapping thus facilitating extraction further magnifying potential DIP high-quality reconstruction. Last, dual U-shape powerful regularizer capture statistics, U-Nets coupled together cross-attention (CAG) module separately achieve spatial feature final generation. CAG incorporate abundant into reconstruction process guide toward more plausible Extensive experiments demonstrate effectiveness our proposed terms quantitative evaluation visual quality. code available at https://github.com/JiaxinLiCAS.

Language: Английский

---

ZMFF: Zero-shot multi-focus image fusion

Information Fusion, Journal Year: 2022, Volume and Issue: 92, P. 127 - 138

Published: Nov. 23, 2022

Language: Английский

---

From Trained to Untrained: A Novel Change Detection Framework Using Randomly Initialized Models With Spatial–Channel Augmentation for Hyperspectral Images

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2023, Volume and Issue: 61, P. 1 - 14

Published: Jan. 1, 2023

Deep learning approaches have been extensively applied to change detection in hyperspectral images (HSIs). However, the majority of them encounter scarcity training samples or rely on complex structures and strategies. Although untrained models proved be effective relief above problems, they were constructed using regular convolutions treated spatial locations channels equally, which are insufficient extract discriminative features lead limited accuracy. Given this, a novel framework randomly initialized with spatial-channel augmentation (RICD) is proposed for HSI this paper. It consists two major modules: 1) an enhanced feature extraction network successive dilation-deformable blocks, can multiscale spatial-spectral over unfixed sampling locations. enlarges field view takes arbitrary neighborhood into consideration, helps increase discriminativeness extracted features; 2) sensitive comparison module integrating selection strategies, exploit context channel importance. magnifies difference between changed pixels unchanged ones emphasizes contribution significant selected features. Despite that convolution operations included RICD, all weights fixed once initialized, indicating RICD work unsupervised manner. Its performance tested three widely used datasets. Quantitative qualitative comparisons several state-of-the-art methods reveal effectiveness method.

Language: Английский

---

Sparse Time–Frequency Analysis of Seismic Data: Sparse Representation to Unrolled Optimization

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2023, Volume and Issue: 61, P. 1 - 10

Published: Jan. 1, 2023

Time-frequency analysis (TFA) is widely used to describe local time-frequency (TF) features of seismic data. Among the commonly TFA tools, sparse (STFA) an excellent one, which can obtain a TF spectrum with good readability. However, many STFA algorithms suffer from expensive calculation time and unavoidable prior knowledge, such as iterative shrinkage-thresholding algorithm (ISTA) reconstruction by separable approximation (SpaRSA). Inspired unrolled its successful applications in signal processing, we propose deep learning-based ISTA algorithm, named network (STFANet). The STFANet contains two parts, i.e., generator module. former learns how transform one-dimensional (1D) large amount unlabelled data into two-dimensional (2D) spectrum, implemented based on proposed dynamic (UIDST) algorithm. Note that UIDST carried out using simplified learning network. latter serves physical constraint model training ensure our obtains accurate actually inverse transform. In this study, traditional short-time Fourier (STFT) utilized To test effectiveness model, apply it 3D post-stack field results show that, compared availably compute better readability, benefits attenuation delineation.

Language: Английский

---

Differentiable Imaging: A New Tool for Computational Optical Imaging

Advanced Physics Research, Journal Year: 2023, Volume and Issue: 2(6)

Published: March 23, 2023

Abstract The field of computational imaging has made significant advancements in recent years, yet it still faces limitations due to the restrictions imposed by traditional techniques. Differentiable programming offers a solution combining strengths classical optimization and deep learning, enabling creation interpretable model‐based neural networks. Through integration physics into modeling process, differentiable imaging, which employs potential overcome challenges posed sparse, incomplete, noisy data. As result, play key role advancing its various applications.

Language: Английский

---

Untrained deep network powered with explicit denoiser for phase recovery in inline holography

Applied Physics Letters, Journal Year: 2023, Volume and Issue: 122(13)

Published: March 27, 2023

Single-shot reconstruction of the inline hologram is highly desirable as a cost-effective and portable imaging modality in resource-constrained environments. However, twin image artifacts, caused by propagation conjugated wavefront with missing phase information, contaminate reconstruction. Existing end-to-end deep learning-based methods require massive training data pairs environmental system stability, which very difficult to achieve. Recently proposed prior (DIP) integrates physical model formation into neural networks without any requirement. process fitting output single measured results interference-related noise. To overcome this problem, we have implemented an untrained network powered explicit regularization denoising (RED), removes images noise Our work demonstrates use alternating directions multipliers method (ADMM) combine DIP RED robust single-shot recovery process. The ADMM, based on variable splitting approach, made it possible plug play different denoisers need differentiation. Experimental show that sparsity-promoting give better over terms signal-to-noise ratio (SNR). Considering computational complexities, conclude total variation denoiser more appropriate for

Language: Английский

---

Fourier Ptychographic Microscopy 10 Years on: A Review

Cells, Journal Year: 2024, Volume and Issue: 13(4), P. 324 - 324

Published: Feb. 10, 2024

Fourier ptychographic microscopy (FPM) emerged as a prominent imaging technique in 2013, attracting significant interest due to its remarkable features such precise phase retrieval, expansive field of view (FOV), and superior resolution. Over the past decade, FPM has become an essential tool microscopy, with applications metrology, scientific research, biomedicine, inspection. This achievement arises from ability effectively address persistent challenge achieving trade-off between FOV resolution systems. It wide range applications, including label-free imaging, drug screening, digital pathology. In this comprehensive review, we present concise overview fundamental principles compare it similar techniques. addition, study on colorization restored photographs enhancing speed FPM. Subsequently, showcase several utilizing previously described technologies, specific focus pathology, three-dimensional imaging. We thoroughly examine benefits challenges associated integrating deep learning To summarize, express our own viewpoints technological progress explore prospective avenues for future developments.

Language: Английский

---

SparseTFNet: A Physically Informed Autoencoder for Sparse Time–Frequency Analysis of Seismic Data

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2022, Volume and Issue: 60, P. 1 - 12

Published: Jan. 1, 2022

The time-frequency (TF) analysis is an effective tool in seismic signal processing. sparsity-based TF transforms have been widely used to obtain high localized representations recent past years. These formulate a sparse representation as inverse optimization problem using simple mathematical models, which are typically based on hand-crafted prior knowledge. Unlike the traditional transforms, supervised deep learning (DL)-based don't require this knowledge and instead use large amount of labeled data set, difficult label for data. In study, bridge gap between DL-based we propose approach physically informed autoencoder model, named SparseTFNet. proposed SparseTFNet includes two modules: convolutional neural networks (CNN)-based encoder representation-based decoder. CNN-based implemented by training absence "ground-truth" representation, can be trained with only traces. short time Fourier transform (STFT) utilized decoder module physical constraint ensure accuracy calculated representation. Finally, after validating model noise-free noisy synthetic traces, applied three-dimensional (3D) offshore results show that has good performance delineation depositional fluvial channels.

Language: Английский

---

Single-shot Fresnel incoherent correlation holography via deep learning based phase-shifting technology

Optics Express, Journal Year: 2023, Volume and Issue: 31(8), P. 12349 - 12349

Published: March 23, 2023

Fresnel incoherent correlation holography (FINCH) realizes non-scanning three-dimension (3D) images using spatial illumination, but it requires phase-shifting technology to remove the disturbance of DC term and twin that appears in reconstruction field, thus increasing complexity experiment limits real-time performance FINCH. Here, we propose a single-shot via deep learning based (FINCH/DLPS) method realize rapid high-precision image only collected interferogram. A network is designed implement operation The trained can conveniently predict two interferograms with phase shift 2/3 π 4/3 from one input Using conventional three-step algorithm, FINCH obtain through back propagation algorithm. Mixed National Institute Standards Technology (MNIST) dataset used verify feasibility proposed experiments. In test MNIST dataset, results demonstrate addition reconstruction, FINCH/DLPS also effectively retain 3D information by calibrating distance case reducing experiment, further indicating superiority method.

Language: Английский

---

Unsupervised crosstalk suppression for self-interference digital holography

Optics Letters, Journal Year: 2025, Volume and Issue: 50(4), P. 1261 - 1261

Published: Jan. 22, 2025

Self-interference digital holography extends the application of to non-coherent imaging fields such as fluorescence and scattered light, providing a new solution, best our knowledge, for wide field 3D low coherence or partially coherent signals. However, cross talk information has always been an important factor limiting resolution this method. The suppression is complex nonlinear problem, deep learning can easily obtain its corresponding model through data-driven methods. in real experiments, it difficult paired datasets complete training. Here, we propose unsupervised method based on cycle-consistent generative adversarial network (CycleGAN) self-interference holography. Through introduction saliency constraint, model, named crosstalk suppressing with neural (CS-UNN), learn mapping between two image domains without requiring training data while avoiding distortions content. Experimental analysis shown that suppress reconstructed images need strategies large number datasets, effective solution technology.

Language: Английский

---