Images obtained under low-light environments generally have color distortions caused by the low intensity of illuminant, its incidence conditions, or camera noises. The color distortions will reduce the performance of various image processing systems which take the low light image as input. Based on assumptions for the scene, many studies on color constancy, trying to find out intrinsic color of objects regardless of colored illuminant, belong to a class of theoretical solutions to the problem of color distortion. However, the preconditions are not practical to low-light images in many cases. That is why the assumptions which are adopted by color constancy methods do not always satisfy phenomena arose in a scene.
To resolve the problem described above, the thesis proposes a color constancy method using the dichromatic reflection model. Because the reflection in the model is defined as either specular reflection or diffused reflection, the bright regions in low-light images are modeled as two factors: the bright regions affected by specular reflection, and the relative dark regions as by diffused reflection. Futhermore, considered in the whole scene, diffused deflection can be regarded as uniform illumination because of affecting all the positions within an image with the same quantity, and specular deflection can be dealt with as non-uniform illumination because of affecting some particular positions with different quantify. For the reason, the proposed method considers gray World(GW) for uniform illumination or diffused deflection, and Multi Scale Retinex with Color Restoration(MSRCR) for non-uniform illumination or specular deflection in order to restore intrinsic color of objects. However, these two methods must be appropriately combined on account of GW having the drawback under non-uniform environments and of MSRCR having the defect under uniform illumination. In this paper, a robust combination is proposed. It gives different weights to two result images, the one from GW and the other from MSRCR, on the basis of the channel intensity of the result image from GW. The advantage of this method is to complement the drawback of each method- GW overcomes the drawback of MSRCR and MSRCR makes up for the deficiency of GW.
The performance of the proposed method was evaluated with lots of synthetic images and with natural images which are acquired in Barnard''s dataset. The synthetic images are used for the evaluation with objectivity and practicality of the proposed method are evaluated using the natural images. As measures for the evaluation of the proposed method, euclid distance, angular error and image constancy were adopted and compared to well-known color constancy methods. Experiment results from the thesis showed that they are generally superior to the existing methods in terms of performance with the all three measures.