SNIPS: Solving Noisy Inverse Problems Stochastically

Authors: Bahjat Kawar, Gregory Vaksman, Michael Elad

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the abilities of the proposed paradigm for image deblurring, super-resolution, and compressive sensing. We demonstrate SNIPS on image deblurring, single image super resolution, and compressive sensing, all of which contain non-negligible noise, and emphasize the high perceptual quality of the results, their diversity, and their relation to the MMSE estimate.
Researcher Affiliation Academia Bahjat Kawar, Gregory Vaksman, Michael Elad Computer Science Department, Technion, Haifa, Israel {bahjat.kawar, grishav, elad}@cs.technion.ac.il
Pseudocode Yes Algorithm 1: SNIPS
Open Source Code No The paper does not provide any statement about releasing source code or a link to a code repository.
Open Datasets Yes In our experiments we use the NCSNv2 [45] network in order to estimate the score function of the prior distribution. Three different NCSNv2 models are used, each trained separately on the training sets of: (i) images of size 64x64 pixels from the Celeb A dataset [27]; (ii) images of size 128x128 pixels from LSUN [56] bedrooms dataset; and (iii) LSUN 128x128 images of towers.
Dataset Splits No The paper mentions "training sets" and "test sets" but does not specify validation splits or detailed percentages for any data splits (training, validation, or test).
Hardware Specification No The paper mentions using a network for experiments but does not provide any specific details about the hardware used (e.g., GPU models, CPU types, memory).
Software Dependencies No The paper mentions using "NCSNv2 [45] network" but does not list any specific software dependencies with version numbers (e.g., Python, PyTorch, specific libraries with versions).
Experiment Setup Yes For image deblurring, we use a uniform 5x5 blur kernel, and an additive white Gaussian noise with σ0 = 0.1 (referring to pixel values in the range [0, 1]). For super resolution, the images are downscaled using a block averaging filter, i.e., each non-overlapping block of pixels in the original image is averaged into one pixel in the low-resolution image. We use blocks of size 2x2 or 4x4 pixels, and assume the low-resolution image to include an additive white Gaussian noise. For compressive sensing, we use three random projection matrices with singular values of 1, that compress the image by 25%, 12.5%, and 6.25%.