[Home]

Conclusion

Camera Analysis

Once the images shoot by the camera are corrected for the camera's gamma conversion,

they are compared to the original images. This comparison is done by calculating the difference between the original image target colors and the final image target colors. This calculation is done with the CompareImagesCenterYes.m script. The target colors from the final images are obtained by averaging together the pixels in the four corner squares as shown in the picture to the right. These differences are then plotted relative to the original image background colors. This results are shown in the following plots which describing the camera's color balancing results. This data in also stored in the centerYesDifferences.mat file.

Red Green Blue

These plots show that the target colors are not dependent on the background colors for the short and middle background wavelengths. During this range the R, G and B values only range ± 60.

The long background wavelengths cause the Red values to dip by almost -100 and the Blue values to rise by over 200. The Green values, however, remain fairly constant with respect to the changes in the background color. This causes a yellow input target color to change to cyan.

Camera Algorithm Simulation

From the analysis of Gray World Algorithm Results and Perfect Reflector Algorithm Results, the Gray world algorithm outputs are closer to the camera outputs comparing to Perfect Reflector algorithm outputs, even they do not match exactly. We draw this conclusion based on:

The overall differences of Gray World algorithm and camera outputs are smaller than the the differences of Perfect Reflector algorithm and camera outputs, which are in [-50, 50] vs. [-170, 100].
As human's intuition and the output data of two Stanford Tower images, the output images of Gray World algorithm are closer to the output images of the camera.

As we said before, each color balancing algorithm is good at some light environments, but not suitable for some other light environments. Using only one algorithm for color balancing limits the quality of the pictures taken by digital cameras, and makes digital cameras not competitive over other products. Therefore, instead of choosing one simple color balancing algorithm, many of the existing cameras use the combination of more than two algorithms or do some optimization for color balancing algorithms. That is why we do not expect the exact match of the camera outputs with algorithm outputs. However, we can see which algorithm we test in this project is closer to the color balancing algorithm used by the camera .

Future Work

The project can be extended in the following respects:

Evaluate other color balancing algorithms, such as color correlation.
Simulate camera color balancing algorithms with combination of two or more algorithms, or with some optimization.

trek@alumni.stanford.edu
lihui@leland.stanford.edu