Color Constancy Research On Low Lighting

Color constancy is taken as the effect of human visual system whereby the perceived color of a surface remains constant despite changes in the spectral composition of the illumination,while the image color will be affected by ambient light.If this perception characteristic can be given to the computer,it can help solve some image processing problems that are highly related to color characteristics.In the past decades of research on color constancy,researchers focused on solving the problem of color constancy in uniform light,while ignoring complex but common scenes such as scenes with low illumination and un-uniform light.There is also a lack of related dataset.To this end,a new low-illumination dataset is established in this paper,and the ground truth is provided.Furthermore,a region selection algorithm based on bright pixels is designed to improve the performance and robustness of some color constancy algorithms on low-illuminance images.In this paper,a color constancy dataset based on night scenes is established,which has the characteristics of low brightness and uneven illumination.The dataset is taken with two different cameras and collects a total of 459 linear images.When shooting at night,the light sources are mostly local light sources,such as streetlights,car lights,and light plates.These local light sources can only light limited areas,as a result,night images have a large number of shadow areas.At the same time,the ISO setting of camera is high in night time,resulting in high noise of the image.The research motivation of the proposed algorithm in this paper comes from an eye movement experiment related to color constancy.Inspired by this experiment,this paper takes brightness information as the core and proposes an algorithm for computational color constancy based on region selection.The proposed algorithm selects areas with low noise that are good for illuminant estimation.Then constraining the reflection characteristics of the selected local areas to calculate multiple light sources at different positions on the image.In the multi-light source estimation problem,the spatial distance and reflection similarity between each pixel and multiple light sources are used to assign light source weights to the entire image,and the estimated light source of the whole image is calculated.Finally,the proposed algorithm is not only verified on self-built low-illuminance dataset,but also tested on public single-light source datasets and multi-light source datasets to prove the general applicability of this method.Experimental results show that performing a region selection step before color constancy algorithm can not only greatly improve the performance of the color constancy algorithm,but also improve the algorithm's robustness.