Normalization

How pixel value normalization works and which preset to choose when uploading a model.

Think of a digital image as a big grid filled with tiny colored dots, commonly called "pixels." Each pixel contains colors - usually a mix of red, green, and blue. These colors have values that range from 0 to 255. A value of 0 means there is none of that color in the pixel (it's totally off), and a value of 255 means that color is shining as brightly as possible.

To make it easier for a computer to analyze and compare different images, the color values can be normalized. One way to do this is by "normalizing" the color values in the image. Normalization is a term for adjusting these values so they fit within a new, consistent range, which helps in comparing images more fairly.

Here's how the normalization formula works:

• normalized_color_value = (original_color_value - mean) / scale;

In this formula:

• original_color_value is the initial value of the color (anywhere from 0 to 255).

• mean is the average of all the color values. Subtracting this mean helps center the color values around zero.

• scale is a number to divide by to keep the values within a new, smaller range. This could be something like the largest color difference or another predefined number.

For example:

• If the average (mean) color value is 100, and the scale is 50:

  • For a pixel with a red color value of 150:

    • Subtract the mean: 150 - 100 = 50

    • Then divide by the scale: 50 / 50 = 1

  • So, the normalized red value would be 1.

This process transforms the original color values to a new scale that's easier for the computer to work with, typically ranging between -1 and 1 or 0 and 1. This is comparable to converting measurements from different units into a single common unit for easier comparison.

In the AI Manager Cloud, set the normalization values as an integer array. For instance, for an RGB image, you might use [123,234,242] for means and [100,232,33] for scales. To leave the original_color_value unchanged, use 0 for mean and 1 for scale, that is, [0,0,0] and [1,1,1] for an RGB input.

The Normalization step in the upload wizard: the Default preset covers most ONNX models.