A mean NIR image (Im) was created for each subject, by taking mean of bands over a spectral range from 750  950 nm. This spans the whole NIR low absorption window. The low absorption window is the range of wavelength from 750 to 950 nm in which main absorbers of skin have low absorption coefficient, which allows the radiations (NIR light) to penetrate deeper in the skin tissues
[10]. These mean NIR images serve as reference images since the contrast per pixel (cpp) for these images is higher than the mean images obtained in other spectral regions. Moreover, the noise, present in each band (of 2 nm) is reduced in the mean image due to its highly random nature. The idea is to find the optimum illumination range within NIR region. There are no hard boundaries between bands and the spectrum seems continuous. Four submean images are created by taking mean of bands in the following ranges:
Im: mean 750950 nm
Image1: mean 750800 nm
Image2: mean 800850 nm
Image3: mean 850900 nm
Image4: mean 900950 nmFrom the hyper spectral dataset, 12 subjects from each skin class were selected randomly. With the data from selected subjects, mean reference image called Im and four submean images, named Image14, were created for each subject of all four classes. Figure
6, Figure
7, Figure
8 and Figure
9 depicts reference images Im and four submean images for a random subject among each skin class. These submean images were then analyzed with the reference of mean reference image Im.
The ultimate users of images are human beings. The most trustworthy way of quality assessment of images is subjective analysis. It is based on human visual system (HVS). The mean opinion score (MOS) is measured by the human viewers. However, this measure is expensive and time consuming
[18]. Furthermore, the MOS is severely affected by the image viewing conditions. In this work objective image quality assessment is chosen due to the following reasons.

1.
Human subjects are unable to distinguish between image quality of all submean images since they look quite similar.

2.
Contrast between veins and skin tissues cannot be determined fairly based on the human visual system.

3.
The acquired NIR images are fed to veins detection algorithm which provide an objective assessment observation.
Objective quality measurement is important for machine vision applications. Mathematical measures are used to measure and compare the image quality w.r.t. reference images. The factors, like mean square error (MSE) and Universal Image Quality Index (Q) are widely used to calculate the quality of images with a reference image which is considered the best image of the scene
[19]. In this work, these two factors are chosen to find out the best range of wavelengths on which one can have good quality image.
Both images (Im and submean images) are converted in to 1D vector before calculating the MSE. For simplicity of notation, we named reference image Im as 'x’ and the sub mean image for which we want to calculate MSE and Q as 'y’.
\mathit{MSE}=\frac{1}{N}{\displaystyle {\sum}_{i=1}^{N}}{\left(\phantom{\rule{0.25em}{0ex}}{x}_{i}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}{y}_{i}\right)}^{2}
(2)
Where 'N’ is the total number of pixels in both images. Furthermore ’xi’ and 'yi’ is the ith pixel in image x and y respectively
[20]. Image x is the reference image and Image y is the one for which we want to calculate the MSE value. The universal image quality index is a measure that is independent of viewing conditions. The range of Q is [1, 1] and is defined by the following equation:
\mathrm{Q}=\frac{{\mathrm{\sigma}}_{\mathrm{xy}}}{{\mathrm{\sigma}}_{\mathrm{x}}{\mathrm{\sigma}}_{\mathrm{y}\phantom{\rule{0.25em}{0ex}}}}\phantom{\rule{0.25em}{0ex}}\times \frac{2\overline{\mathrm{x}}\phantom{\rule{0.25em}{0ex}}\overline{\mathrm{y}}}{{\left(\overline{\mathrm{x}}\right)}^{2}+{\left(\overline{\mathrm{y}}\right)}^{2}}\phantom{\rule{0.25em}{0ex}}\times \frac{2{\mathrm{\sigma}}_{\mathrm{x}}{\mathrm{\sigma}}_{\mathrm{y}\phantom{\rule{0.25em}{0ex}}}}{{\mathrm{\sigma}}_{\mathrm{x}}^{2}+{\mathrm{\sigma}}_{\mathrm{y}\phantom{\rule{0.25em}{0ex}}}^{2}}
(3)
Where
\overline{x}=\frac{1}{N1}{\displaystyle {\sum}_{i=1}^{N}{x}_{i}}\cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \overline{y}=\frac{1}{N1}{\displaystyle {\sum}_{i=1}^{N}{y}_{i}}
{\sigma}_{x}^{2}=\frac{1}{N1}{\displaystyle \sum _{i=1}^{N}}{\left(\phantom{\rule{0.25em}{0ex}}{x}_{i}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\overline{x}\right)}^{2},\phantom{\rule{0.25em}{0ex}}{\sigma}_{y}^{2}=\frac{1}{N1}{\displaystyle \sum _{i=1}^{N}}{\left(\phantom{\rule{0.25em}{0ex}}{y}_{i}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\overline{y}\right)}^{2}
{\sigma}_{x}{\sigma}_{y}\phantom{\rule{0.25em}{0ex}}=\frac{1}{N\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}1}{\displaystyle \sum _{i=1}^{N}}\left({x}_{i}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\overline{x}\right)\left({y}_{i}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\overline{y}\right)
There are three components of Q in Eq.3; the first component is the coefficient of correlation between images, the reference image and the one whose quality factor is being measured. With the second component, the relation of luminance of both images is measured. Third component measures the similarity of contrast of both images.In Figure
10 the mean value of MSE calculated for four submean images for all 12 subjects of each class is plotted. In this plot it can be observed that the MSE for the submean image (Image2), which was formed by taking mean in the range of 800 850 nm bands, has the lowest value of MSE for all skin classes. The 4th image (Image4) which was formed by taking mean in the range of 900 950 nm bands has the highest MSE for all skin classes.The Q factor is plotted in Figure
11. In this plot the value of Q factor is higher for Image2 as compared to other 3 images except in case of fair skin. In that case the Q value for Image1 is slightly higher than Image2, but the difference is not that big which can lead to any conclusive remarks. The overall results are consistent with the MSE values.
Through this work, it is determined that the image (image2) which was made with the mean of bands in range 800850 nm from multispectral data has the best quality. The measure like MSE, PSNR and universal image quality (Q) is found best for this range. Keeping in mind the image acquisition setup, it is concluded that the best quality image is obtained in the spectral range of 800 – 850 nm. These findings will serve the basis of optimum illumination selection for a NIR system for subcutaneous veins localization. For this system, our choice will be the LEDs with a central wavelength lying within 800850 nm range. With this optimized illumination it is anticipated that better quality and high contrast NIR images can be obtained.