In this paper, we introduce convolutional proximal neural networks (cPNNs), which are by construction averaged operators. For filters with full length, we propose a stochastic gradient descent algorithm on a submanifo...
详细信息
In this paper, we introduce convolutional proximal neural networks (cPNNs), which are by construction averaged operators. For filters with full length, we propose a stochastic gradient descent algorithm on a submanifold of the Stiefel manifold to train cPNNs. In case of filters with limited length, we design algorithms for minimizing functionals that approximate the orthogonality constraints imposed on the operators by penalizing the least squares distance to the identity operator. Then, we investigate how scaled cPNNs with a prescribed Lipschitz constant can be used for denoising signals and images, where the achieved quality depends on the Lipschitz constant. Finally, we apply cPNN based denoisers within a plug-and-play framework and provide convergence results for the corresponding PnP forward-backward splitting algorithm based on an oracle construction. (c) 2021 Elsevier Inc. All rights reserved.
For magnetic resonance imaging (MRI), recently proposed "plug-and-play" (PnP) image recovery algorithms have shown remarkable performance. These PnP algorithms are similar to traditional iterative algorithms...
详细信息
ISBN:
(纸本)9780738131269
For magnetic resonance imaging (MRI), recently proposed "plug-and-play" (PnP) image recovery algorithms have shown remarkable performance. These PnP algorithms are similar to traditional iterative algorithms like FISTA, ADMM, or primal-dual splitting (PDS), but differ in that the proximal update is replaced by a call to an application-specific image denoiser, such as BM3D or DnCNN. The fixed-points of PnP algorithms depend upon an algorithmic stepsize parameter, however, which must be tuned for optimal performance. In this work, we propose a fast and robust auto-tuning PnP-PDS algorithm that exploits knowledge of the measurement-noise variance that is available from a pre-scan in MRI. Experimental results show that our algorithm converges very close to genie-tuned performance, and does so significantly faster than existing autotuning approaches.
In this paper, we present the Pipeline for Observational Data Processing, Analysis, and Collaboration (PODPAC) software. PODPAC is an open-source Python library designed to enable widespread exploitation of NASA earth...
详细信息
In this paper, we present the Pipeline for Observational Data Processing, Analysis, and Collaboration (PODPAC) software. PODPAC is an open-source Python library designed to enable widespread exploitation of NASA earth science data by enabling multi-scale and multi-windowed access, exploration, and integration of available earth science datasets to support analysis and analytics;automatic accounting for geospatial data formats, projections, and resolutions;simplified implementation and parallelization of geospatial data processing routines;standardized sharing of data and algorithms;and seamless transition of algorithms and data products from local development to distributed, serverless processing on commercial cloud computing environments. We describe the key elements of PODPAC's architecture, includingNodesfor unified encapsulation of disparate scientific data sources;algorithmsfor plug-and-play processing and harmonization of multiple data sourceNodes;andLambda functionsfor serverless execution and sharing of new data products via the cloud. We provide an overview of our open-source code implementation and testing process for development and deployment of PODPAC. We describe our interactive, JupyterLab-based end-user documentation including quick-start examples and detailed use case studies. We conclude with examples of PODPAC's application to: encapsulate data sources available on Amazon Web Services (AWS) Open Data repository;harmonize processing of multiple earth science data sets for downscaling of NASA Soil Moisture Active Passive (SMAP) soil moisture data;and deploy a serverless SMAP-based drought monitoring application for use access from mobile devices. We postulate that PODPAC will also be an effective tool for wrangling and standardizing massive earth science data sets for use in model training for machine learning applications.
Cardiac magnetic resonance imaging (CMR) is a noninvasive imaging modality that provides a comprehensive evaluation of the cardiovascular system. The clinical utility of CMR is hampered by long acquisition times, howe...
详细信息
ISBN:
(纸本)9781538693308
Cardiac magnetic resonance imaging (CMR) is a noninvasive imaging modality that provides a comprehensive evaluation of the cardiovascular system. The clinical utility of CMR is hampered by long acquisition times, however. In this work, we propose and validate a plug-and-play (PnP) method for CMR reconstruction from undersampled multi-coil data. To fully exploit the rich image structure inherent in CMR, we pair the PnP framework with a deep learning (DL)-based denoiser that is trained using spatiotemporal patches from high-quality, breath-held cardiac cine images. The resulting "PnP-DL" method iterates over data consistency and denoising subroutines. We compare the reconstruction performance of PnP-DL to that of compressed sensing (CS) using eight breath-held and ten real-time (RT) free-breathing cardiac cine datasets. We find that, for breath-held datasets, PnP-DL offers more than one dB advantage over commonly used CS methods. For RT free-breathing datasets, where ground truth is not available, PnP-DL receives higher scores in qualitative evaluation. The results highlight the potential of PnP-DL to accelerate RT CMR.
The expectation consistent (EC) approximation framework is a state-of-the-art approach for solving (generalized) linear inverse problems with high-dimensional random forward operators and i.i.d. signal priors. In imag...
详细信息
ISBN:
(纸本)9798350325744
The expectation consistent (EC) approximation framework is a state-of-the-art approach for solving (generalized) linear inverse problems with high-dimensional random forward operators and i.i.d. signal priors. In image inverse problems, however, both the forward operator and image pixels are structured, which plagues traditional EC implementations. In this work, we propose a novel incarnation of EC that exploits deep neural networks to handle structured operators and signals. For phase-retrieval, we propose a simplified variant called "deepECpr" that reduces to iterative denoising. In experiments recovering natural images from phaseless, shot-noise corrupted, coded-diffraction-pattern measurements, we observe accuracy surpassing the state-of-the-art prDeep (Metzler et al., 2018) and Diffusion Posterior Sampling (Chung et al., 2023) approaches with two-orders-of-magnitude complexity reduction.
暂无评论