Quantum image rain removal: second-order photon number fluctuation correlations in the time domain

TL;DR

This paper introduces a quantum optical method leveraging second-order photon number fluctuation correlations to effectively remove rain-induced noise from images, outperforming some deep learning techniques in short measurement scenarios.

Contribution

It presents a novel quantum measurement approach for rain removal based on second-order photon correlation, offering a new front-end processing technique for image enhancement.

Findings

Quantum correlation properties distinguish rain photons from non-rain photons.

The method outperforms deep learning in short measurement times.

Experimental results validate the effectiveness of quantum rain removal.

Abstract

Falling raindrops are usually considered purely negative factors for traditional optical imaging because they generate not only rain streaks but also rain fog, resulting in a decrease in the visual quality of images. However, this work demonstrates that the image degradation caused by falling raindrops can be eliminated by the raindrops themselves. The temporal second-order correlation properties of the photon number fluctuation introduced by falling raindrops has a remarkable attribute: the rain streak photons and rain fog photons result in the absence of a stable second-order photon number correlation, while this stable correlation exists for photons that do not interact with raindrops. This fundamental difference indicates that the noise caused by falling raindrops can be eliminated by measuring the second-order photon number fluctuation correlation in the time domain. The simulation and experimental results demonstrate that the rain removal effect of this method is even better than that of deep learning methods when the integration time of each measurement event is short. This high-efficient quantum rain removal method can be used independently or integrated into deep learning algorithms to provide front-end processing and high-quality materials for deep learning.