Convert colour to grey

Converting a colour picture to grey replaces the three numbers of each pixel with one number for its brightness. Most vision methods start from grey, because one number per pixel is easier to compare than three.

The skill is doing it so that the grey picture still looks right to a person.

A colour pixel is three numbers, so it is hard to ask whether one pixel is brighter than another. A grey pixel is one number, so a comparison is easy.

Edges, thresholds and counting all work on grey levels. We convert once, then reuse the result.

The simple way is to average the three channels. It looks wrong. The eye is most sensitive to green and least to blue. Equal weights make a green field look too dark and a blue sky too bright.

Luminance weights fix this. The usual ones are 0.299 for red, 0.587 for green and 0.114 for blue. They add up to 1.

docker exec jupyter python3 -c "from PIL import Image; print([Image.new('RGB', (1, 1), c).convert('L').getpixel((0, 0)) for c in [(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 255)]])"

The Run above makes four one-pixel pictures and converts each with convert("L"). It prints [76, 150, 29, 255]: red 76, green 150, blue 29 and white 255.

Those match the table, so Pillow uses the same weights. White stays at 255, because the weights add up to 1.

red = 200
green = 40
blue = 40

average = (red + green + blue) // 3
weighted = (299 * red + 587 * green + 114 * blue + 500) // 1000

print("simple average:", average)
print("weighted grey:", weighted)
# Change the pixel to red 40, green 200, blue 40 and press Run.
# Which one gets brighter, and why?

docker exec jupyter python3 -c "from PIL import Image; p = Image.open('images/shapes.png'); print(p.getpixel((50, 40)), p.convert('L').getpixel((50, 40)), p.getpixel((150, 140)), p.convert('L').getpixel((150, 140)), p.getpixel((150, 45)), p.convert('L').getpixel((150, 45)))"

The Run reads one pixel from three shapes of shapes.png. The red shape is (200, 40, 40) and becomes 88. The blue one is (40, 70, 200) and becomes 76. The green one is (40, 160, 70) and becomes 114.

Three clear colours, but the greys sit close together. Grey throws the colour away, so a red and a blue shape can be hard to tell apart.

Sometimes one channel is the useful one. A red shape on a white page is easier to find in the green channel, where it is dark. In the red channel it is as bright as the page.

In NumPy, a[:, :, 0] is the red channel, a[:, :, 1] the green and a[:, :, 2] the blue.

In a notebook cell, load shapes.png as an array and print the mean of each channel with a[:, :, 0].mean() and the same for 1 and 2.

The page is mostly white, so the three means are close. Which shape would be hardest to find in grey, and why?