The Battle of the Algorithms: Testing and User Interface

To save time when color balancing, we tried a variant of our algorithms that first blurs (or lowpass filters) the images before calculating the transformation coefficients in the color balancing algorithms. If the results of calculating the transformation coefficients of the color balancing on a blurred version of an image are the same as calculating the transformation coefficients of the original are the same, we can potentially save computation time by downsampling the original image before calculating the color balancing coefficients. The smaller area image will allow for faster computation, something that will become increasingly important as the sizes of digital camera images continues to increase. The time savings are not particularly significant for many simple algorithms, including Grayworld, but they can potentially be quite large for more complicated color balancing solutions.

We tested this procedure on our Grayworld algorithm and our Mean/Standard Deviation algorithm. We found that the MSE between the resulting image that was color balanced using our original algorithms and those that were first severely blurred, and then color balanced was very nearly zero. This result seems to indicate that low-pass filtering and downsampling could potentially be very beneficial for certain algorithms. However, without testing, we know that it will also fail for some algorithms under certain circumstances. For instance, with the White world algorithm, the blurring may cause the brightest parts of the image to decrease in intensity as they are averaged with darker spots. This intensity decrease will lead to a larger transformation in the original image, thus giving a different result. In conclusion, we can say that blurring the images before processing could potentially be a time saver for some algorithms and for most scenes, but it is not a robust solution that will work under all circumstances.

Region-based Standard Deviation Adjustment

Given the poor results we got for the high contrast images, we implemented an algorithm that adjusts the standard deviation differently in different regions of the image. Basically, this algorithm identifies two regions in an image: one bright region and one dark. It then does a region-based standard deviation adjustment to avoid pushing bright regions into saturation, or dark regions to zero.