Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
NPN: Non-Linear Projections of the Null-Space for Imaging Inverse Problems
Authors: Roman Jacome, Romario Gualdrón-Hurtado, León Suárez-Rodríguez, Henry Arguello
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results across diverse sensing matrices demonstrate that NPN priors consistently enhance reconstruction fidelity in various imaging inverse problems, such as compressive sensing, deblurring, super-resolution, computed tomography, and magnetic resonance imaging, with plug-and-play methods, unrolling networks, deep image prior, and diffusion models. 5 Experiments The proposed NPN regularization was evaluated in five imaging inverse problems: compressed sensing, super-resolution, computed tomography, single coil MRI, and deblurring. |
| Researcher Affiliation | Academia | Department of Electrical, Electronics, and Telecommunications Engineering Department of Systems Engineering and Informatics Universidad Industrial de Santander, Colombia, 680002 EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 GENERATE ORTHONORMAL ROWS TO H VIA QR DECOMPOSITION |
| Open Source Code | Yes | The code will be uploaded in the supplementary material with sufficient instructions to reproduce the results of the paper. 1github.com/yromariogh/NPN |
| Open Datasets | Yes | Compressed Sensing: The single-pixel camera (SPC) is used along with the CIFAR-10 dataset [28]... MRI: We employed the fast MRI knee single-coil MRI dataset [27]... Deblurring and Super-Resolution: For these experiments, we used the Celeb A [34] dataset... Computed Tomography: ...Lo Do Pa B-CT dataset [30]... Performance in data-driven models and dataset generalization In Table 2 we report PSNR (d B) for both Pn P and unrolling solvers on CIFAR-10 [28] (in-distribution) and STL10 [13] (out-of-distribution)... Deep Image Prior We train G with the Places365 dataset [64]... |
| Dataset Splits | Yes | Compressed Sensing: The single-pixel camera (SPC) is used along with the CIFAR-10 dataset [28], with 50, 000 images for training and 10, 000 for testing. All images were resized to 32 32. ... MRI: We employed the fast MRI knee single-coil MRI dataset [27], which consists of 900 training images and 73 test images of knee MRIs of 320 320. The training set was split into 810 images for training and 90 for validation, and all images were resized to 256 256. ... Deblurring and Super-Resolution: For these experiments, we used the Celeb A [34] dataset resized to 128 128, using 8000 images for training and 2000 for testing. ... Computed Tomography: ...The Lo Do Pa B-CT dataset [30] was resized to 256 256 and used for training; in testing, we used 10 test set slices. ... Deep Image Prior We train G with the Places365 dataset [64], where we used 28.000 images for training and 7000 images for testing. All images were resized to 128 128. |
| Hardware Specification | Yes | All simulations were performed on an NVIDIA RTX 4090 GPU, with code in 1. |
| Software Dependencies | No | The method was implemented using the Py Torch framework. |
| Experiment Setup | Yes | Compressed Sensing: ... The Adam [26] optimizer was used with a learning rate of 5 10 4. ... MRI: ... trained it for 60 epochs with a learning rate of 1 10 4, using the Adam W optimizer [35] with a weight decay of 1 10 2 and a batch size of 4. ... Deblurring and Super-Resolution: ... we employed a U-Net architecture for the network G. In this case, we used the Adam optimizer with a learning rate of 1 10 3 and a batch size of 32. ... Computed Tomography: We train the DM for 1000 epochs with batch size 4 using the Adam W optimizer and learning rate 3 10 4. ... G is a U-Net which was trained for 100 epochs with a learning rate 3 10 4 and a batch size of 4 using Adam W. ... Deep Image Prior ... The network K was trained following (15) using the Adam optimizer with a learning rate of 1e 3 for 1000 iterations. |