High SNR Processing for Low-Light Images

This image enhancement on a single image for better visibility and higher SNR (signal-to-noise ratio) is an under-constrained problem, which becomes more challenging as this image is under the low-light condition. In order to solve the above problems, we firstly analyze the reason that classical Retinex enhancement result has low SNR due to ignoring the influence of noise in reflectance. To this end, combined with the actual illumination-reflectance prior, a new convex optimization function is proposed to estimate noise-suppressed/detail-preserved reflectance and spatial piece-wise smoothed illumination simultaneously. Especially, although the objective of the proposed method is a hybrid non-smooth problem, by subtly decomposition of the objective, we can solve it effectively. The final low-light enhancement result with pleasant visual performance and high SNR can be effectively obtained according to recompose the adjusting illumination and reflectance. Qualitative and quantitative experiments have illustrated that the proposed method has a higher SNR and more pleasing visual effect than state-of-the-art image enhanced techniques under low-light conditions.

I. Introduction

This Low-light images are of low signal-to-noise ratio (SNR), i.e, low visibility and high-level noise. The processing of these images is highly desirable in both providing pleasing visualization for humans and uncovering details for machine vision applications. To this end, many techniques on single low-light image have been proposed which can be divided into four categories: 1) traditional image enhancement [1], [2]; 2) image decomposition [3], [4]; and 3) high-dynamic-range (HDR) imaging [5], [6]; 4) deep learning enhancement [7].