Computer Vision: Quantize RGB for Faster Learning and Better Generalization
There is information missing from the standard RGB color coded system. RGB is an intensity based system with (0,0,0) indicating no light - black - and (255,255,255) full intensity light from all three channels - white. There is a physically based quantity that can be easily added to provide more information -> Energy. The first step is to quantize the RGB system.
Normalize the RGB values (like usual) with the max value, 255, such that each channel has range [0.00,1.00]. Now, choose the quanta size depending on how many unique colors you want to map. Here, I choose 1 quanta equals 0.01 RGB intensity units. This gives [0,100] quanta per channel which works out to 101^3 = 1,030,301 unique colors. Our RGB intensity values are now in units of a quantity analogous to photons.
Choose the quanta energy per channel by taking the energies of characteristic red, green, and blue photons, for example:
🔴 Red photon energy: 1.77 eV (~700nm)
🟢 Green photon energy: 2.25 eV (~550nm)
🔵 Blue photon energy: 2.76 eV (~450nm)
Darkness is 0eV. White light is 678.0 eV with these choices. Calculate the energy at each quantized level of RGB and scale by white light energy. This gives you a map to the RGB system at 0.01 increments per channel to assign the 4th channel.
The 4th channel from above is proportional to the photon energy flux a light receptor such as an eye or sensor would experience in the absorption event. This will speed up your training. Godspeed.
Choose 1/255 as the quanta to map the entire 16,777,216 possible RGB colors.