Abstract

A method of training neural networks for image enhancement without using reference data is proposed based on the assumption of similarity of signals and independence of noise components in spatially proximal image pixels. This approach enables the formation of the training dataset from each frame of a video sequence by decimation into even and odd rows and columns. The training of image restoration is possible considering the markers of dynamic properties of objects in the image. The efficiency and limitations of the proposed method are studied. Its performance is evaluated using a database of images obtained at a low light intensity.

© 2021 Optical Society of America

Stabilization of turbulence-degraded video using patch-based reference frame

Fouzia Nawreen, Kalyan Kumar Halder, Murat Tahtali, and Sreenatha G. Anavatti
Opt. Continuum 2(12) 2484-2499 (2023)

Single-pixel image reconstruction from experimental data using neural networks

Antonio Lorente Mur, Pierre Leclerc, Françoise Peyrin, and Nicolas Ducros
Opt. Express 29(11) 17097-17110 (2021)

Phase-only hologram generated by a convolutional neural network trained using low-frequency mixed noise

Xi Wang, Xinlei Liu, Tao Jing, Pei Li, Xiaoyu Jiang, Qiang Liu, and Xingpeng Yan
Opt. Express 30(20) 35189-35201 (2022)

Single photon compressive imaging with enhanced quality using an untrained neural network

Yuhan Wang and Lingbao Kong
J. Opt. Soc. Am. A 40(12) 2240-2248 (2023)

Mid-fusion of road scene polarization images on pretrained RGB neural networks

Khalid Omer and Meredith Kupinski
J. Opt. Soc. Am. A 38(4) 515-525 (2021)