In NN based approaches, a NN is trained to learn the
relationships between the colors in the image and the expected
illuminant. The advantage of using NNs is that there
are no explicit assumptions regarding the image content as
in Gray World or white patch methods. Nayak and Chaudari
[92] used a NN for color constancy in tracking human
palm. In their approach, a NN (2 ∗ 20 ∗ 3) is trained using
a back propagation to directly learn the illuminant parameters.
The inputs to the NN are RGB components of the
skin pixels, while the output of the network is the expected
canonical RGB components. The NN is trained on a set of
images containing human palm under varying illumination
conditions. The results reported suggest that NN adapts to
the illuminant parameters and NN adapted palm images can
be tracked precisely in a variety of cluttered backgrounds
and varying illuminations.
Kakumanu et al. [93,103] also used NN for color constancy.
The proposed three layered NN (1600 ∗ 48 ∗ 8 ∗ 2)
directly estimates the illuminants so as to bring the skin
color to gray. The input to the NN is an rg histogram and
the output of the network is the expected illuminant of the
skin in rg space. The NN is trained on a dataset of 255
images and tested on 71 images, the images representing a
wide range of illuminations both indoor and outdoor, different
backgrounds and non-white light sources. A simple
thresholding technique is used to detect skin from these NN
color corrected images.