SelfMixed: Self-supervised mixed noise attenuation for distributed acoustic sensing data DOI
Zitai Xu, Bangyu Wu, Yisi Luo

et al.

Geophysics, Journal Year: 2024, Volume and Issue: 89(5), P. V415 - V436

Published: May 29, 2024

Distributed acoustic sensing (DAS) is an emerging data acquisition technique known for its high density, cost effectiveness, and environmental friendliness, making it a technology with significant future application potential in many fields. However, DAS signals are often contaminated by various types of noise, such as high-frequency, high-amplitude erratic, horizontal their processing challenging. Therefore, crucial to leverage the physical characteristics these diverse noise effectively attenuate them. In this work, we develop SelfMixed, novel self-supervised learning method mixed suppression data. We fully exploit different introduce characteristic-based training strategy. Specifically, use [Formula: see text] norm characterize random erratic smoothness vertical nonsmoothness noise. addition, blind-spot-based strategy denoising, relying solely on observed noisy To more also Fourier transform-based parameterization method. By combining deep priors our attenuates complex field Extensive experiments synthetic from geographic scenarios validate superiority SelfMixed over seven state-of-the-art denoising approaches.

Language: Английский

Exploring the values underlying machine learning research in medical image analysis DOI Creative Commons
John S. H. Baxter, Roy Eagleson

Medical Image Analysis, Journal Year: 2025, Volume and Issue: 102, P. 103494 - 103494

Published: Feb. 25, 2025

Language: Английский

Citations

0
Machine Learning Algorithms for Applications in Materials Science I DOI
Azizeh Abdolmaleki, Fereshteh Shiri, Shahin Ahmadi

et al.

Challenges and advances in computational chemistry and physics, Journal Year: 2025, Volume and Issue: unknown, P. 191 - 214

Published: Jan. 1, 2025

Language: Английский

Citations

0
Deep Random Projector: Accelerated Deep Image Prior DOI
Taihui Li,

Hengkang Wang,

Zhong Zhuang

et al.

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown

Published: June 1, 2023

Deep image prior (DIP) has shown great promise in tackling a variety of restoration (IR) and general visual inverse problems, needing no training data. However, the resulting optimization process is often very slow, inevitably hindering DIP's practical usage for time-sensitive scenarios. In this paper, we focus on IR, propose two crucial modifications to DIP that help achieve substantial speedup: 1) optimizing seed while freezing randomly-initialized network weights, 2) reducing depth. addition, reintroduce explicit priors, such as sparse gradient prior-encoded by total-variation regularization, preserve peak performance. We evaluate proposed method three IR tasks, including denoising, super-resolution, inpainting, against original variants, well competing metaDIP uses meta-learning learn good initializers with extra Our clear winner obtaining competitive quality minimal amount time. code available at https://github.com/sun-umn/Deep-Random-Projector.

Language: Английский

Citations

9
Single-shot deep-learning based 3D imaging of Fresnel incoherent correlation holography DOI

Qinnan Zhang,

Tao Huang, Jiaosheng Li

et al.

Optics and Lasers in Engineering, Journal Year: 2023, Volume and Issue: 172, P. 107869 - 107869

Published: Sept. 29, 2023

Language: Английский

Citations

9
Intelligent Inverse Designs of Impedance Matching Circuits With Generative Adversarial Network DOI
Jingwei Zhang, Zhun Wei, Kai Kang

et al.

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, Journal Year: 2024, Volume and Issue: 43(10), P. 3171 - 3183

Published: April 9, 2024

Impedance matching circuits (IMCs) are crucial modules in radio frequency (RF) front-end components, devices, and systems, affecting the performance of whole systems. However, design process IMCs has to require intense manual interventions with high computational costs. To alleviate this problem, a novel scheme for inversely designing is presented work based on neural network technology. Such IMC inverse framework consists two mapping-based deep networks (DNNs). The first one an untrained generative adversarial (GAN) that maps from requirements regularized S-parameters curves. second inversion target impedance designed circuit parameters. With cascaded GAN network, efficient method IMC-based filtering antenna introduced, which takes about 1/17 time compared traditional EM-based optimization methods. Further, three power amplifiers (PAs) multiple inversely-designed proposed framework. In experimental demonstration, elaborate prototypes fabricated measured, where measured results fully satisfy demand performance.

Language: Английский

Citations

3
Incrementally Adapting Pretrained Model Using Network Prior for Multi-Focus Image Fusion DOI
Xingyu Hu, Junjun Jiang, Chenyang Wang

et al.

IEEE Transactions on Image Processing, Journal Year: 2024, Volume and Issue: 33, P. 3950 - 3963

Published: Jan. 1, 2024

Multi-focus image fusion can fuse the clear parts of two or more source images captured at same scene with different focal lengths into an all-in-focus image. On one hand, previous supervised learning-based multi-focus methods relying on synthetic datasets have a distribution shift real scenarios. other unsupervised well adapt to observed but lack general knowledge defocus blur that be learned from paired data. To avoid problems existing methods, this paper presents novel model by considering both brought pretrained backbone and extrinsic priors optimized specific testing sample improve performance fusion. specific, Incremental Network Prior Adaptation (INPA) framework is proposed incrementally integrate features extracted strong baselines tiny prior network (6.9% parameters network) boost for test samples. We evaluate our method real-world public (Lytro, MFI-WHU, Real-MFF) show outperforms learning based methods.

Language: Английский

Citations

3
Multilinear Kernel Regression and Imputation via Manifold Learning DOI Creative Commons
Duc Thien Nguyen, Konstantinos Slavakis

IEEE Open Journal of Signal Processing, Journal Year: 2024, Volume and Issue: 5, P. 1073 - 1088

Published: Jan. 1, 2024

Language: Английский

Citations

3
Multitasking optimization for the imaging problem in electrical capacitance tomography DOI
Jing Lei,

Qibin Liu

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 257, P. 125105 - 125105

Published: Aug. 20, 2024

Language: Английский

Citations

3
A Federated Deep Unrolling Method for Lidar Super-Resolution: Benefits in SLAM DOI Creative Commons
Αλέξανδρος Γκίλλας, Aris S. Lalos, Evangelos Markakis

et al.

IEEE Transactions on Intelligent Vehicles, Journal Year: 2023, Volume and Issue: 9(1), P. 199 - 215

Published: Nov. 9, 2023

In this paper, we propose a novel federated deep unrolling method for increasing the accuracy of Lidar Super resolution. The proposed framework not only offers notable improvements in Lidar-based SLAM methodologies but also provides solution to significant cost associated with high-resolution sensors. Particularly, our can be adopted by number vehicles coordinated server towards learning regularizer - neural network capturing dependencies data. To tackle adaptive optimization problem effectively, initially framework, converting into well-justified architecture. learnable parameters architecture are directly derived from problem, thus resulting an explainable Further, extend capabilities technique incorporating strategy. Our model employs innovative Adapt-then-Combine strategy, where each vehicle optimizes its and, subsequently, their regularizers combined formulate robust global regularizer, equipped handle diverse environmental conditions. Through extensive numerical evaluations on real-world based applications, demonstrates superior performance along reduction trainable parameters, 99.75% fewer compared state art lidar super-resolution networks. Essentially, study is first initiative combine learning, showcasing efficient, and data-secure approach automotive applications. source code found at: https://github.com/alexandrosgk/Federated-Deep-Unrolling-Lidar-Super-resolution-SLAM.git .

Language: Английский

Citations

8
Unsupervised reconstruction with a registered time-unsheared image constraint for compressed ultrafast photography DOI Creative Commons
Haoyu Zhou, Yan Song, Zhiming Yao

et al.

Optics Express, Journal Year: 2024, Volume and Issue: 32(9), P. 16333 - 16333

Published: April 3, 2024

Compressed ultrafast photography (CUP) is a computational imaging technology capable of capturing transient scenes in picosecond scale with sequence depth hundreds frames. Since the inverse problem CUP an ill-posed problem, it challenging to further improve reconstruction quality under condition high noise level and compression ratio. In addition, there are many articles adding external charge-coupled device (CCD) camera system form time-unsheared view because added constraint can images. However, since images collected by different cameras, slight affine transformation may have great impacts on quality. Here, we propose algorithm that combines image unsupervised neural networks. Image registration network also introduced into framework learn parameters input The proposed effectively utilizes implicit prior as well extra hardware information brought view. Combined network, this joint learning model enables our reconstructed without training datasets. simulation experiment results demonstrate application prospect event capture.

Language: Английский

Citations

2