I’m developing a Python script to rank images based on "blueness" for a colorblind accessibility project (I'm colorblind myself). The goal is to prioritize images with large blue areas over those with tiny intense blue spots. Despite adjustments, non-blue images (e.g., grays/whites) are ranking highly due to residual blue channel values.

Problem

The script calculates a combined score (coverage + intensity) but incorrectly ranks images like image20.png (0% coverage, 25% intensity) above truly blue images.

image20.png

Key Requirements

1. Coverage: Percentage of pixels where blue is dominant and sufficiently bright.
2. Intensity: Average blue value only for qualifying pixels.
3. Score: 0.9 * coverage + 0.1 * intensity to prioritize coverage.

Current Code

from PIL import Image
import numpy as np

def average_blue_intensity(image_path):
    img = Image.open(image_path).convert('RGB')
    img_array = np.array(img).astype(np.int16)
    red, green, blue = img_array[:, :, 0], img_array[:, :, 1], img_array[:, :, 2]
    mask = (blue > red + 40) & (blue > green + 40) & (blue > 100)
    return np.mean(blue[mask]) / 255 if np.any(mask) else 0

def blue_coverage(image_path):
    # ... (same mask as above)
    return np.mean(mask)

# Combined score calculation
combined_scores = {path: 0.9 * coverage + 0.1 * intensity ...}

What I’ve Tried

1. Stricter Thresholds: Blue must exceed red/green by 40 units and be >100 in brightness.

2. Debugging Pixels: Confirmed image20.png has no true blue pixels (e.g., grays like (100,100,100)).

3. Visualizing Masks: Generated debug images to verify mask accuracy.

Unexpected Behavior

• image20.png (no blue) scores higher than image127.png (5% coverage).

• Full output:

  1. image20.png - Score: 0.0999 (Coverage: 0.00%, Intensity: 24.97%)  
  2. image127.png - Score: 0.0964 (Coverage: 5.23%, Intensity: 16.26%)  
  

Question

How can I refine the algorithm to exclude non-blue pixels (e.g., grays/whites) and ensure the score truly reflects blue dominance?

I’m developing a Python script to rank images based on "blueness" for a colorblind accessibility project (I'm colorblind myself). The goal is to prioritize images with large blue areas over those with tiny intense blue spots. Despite adjustments, non-blue images (e.g., grays/whites) are ranking highly due to residual blue channel values.

Problem

The script calculates a combined score (coverage + intensity) but incorrectly ranks images like image20.png (0% coverage, 25% intensity) above truly blue images.

image20.png

Key Requirements

1. Coverage: Percentage of pixels where blue is dominant and sufficiently bright.
2. Intensity: Average blue value only for qualifying pixels.
3. Score: 0.9 * coverage + 0.1 * intensity to prioritize coverage.

Current Code

from PIL import Image
import numpy as np

def average_blue_intensity(image_path):
    img = Image.open(image_path).convert('RGB')
    img_array = np.array(img).astype(np.int16)
    red, green, blue = img_array[:, :, 0], img_array[:, :, 1], img_array[:, :, 2]
    mask = (blue > red + 40) & (blue > green + 40) & (blue > 100)
    return np.mean(blue[mask]) / 255 if np.any(mask) else 0

def blue_coverage(image_path):
    # ... (same mask as above)
    return np.mean(mask)

# Combined score calculation
combined_scores = {path: 0.9 * coverage + 0.1 * intensity ...}

What I’ve Tried

1. Stricter Thresholds: Blue must exceed red/green by 40 units and be >100 in brightness.

2. Debugging Pixels: Confirmed image20.png has no true blue pixels (e.g., grays like (100,100,100)).

3. Visualizing Masks: Generated debug images to verify mask accuracy.

Unexpected Behavior

• image20.png (no blue) scores higher than image127.png (5% coverage).

• Full output:

  1. image20.png - Score: 0.0999 (Coverage: 0.00%, Intensity: 24.97%)  
  2. image127.png - Score: 0.0964 (Coverage: 5.23%, Intensity: 16.26%)  
  

Question

How can I refine the algorithm to exclude non-blue pixels (e.g., grays/whites) and ensure the score truly reflects blue dominance?

• python
• image
• color-blindness

1 Answer 1

You are evaluating an image the way the camera sees it, in RGB space.

Consider switching to something closer to human perception, such as HSV space.

A pypi library can help with that.

$ pip install nice-colorsys
$
$ python -c 'from nice_colorsys import hsv; print("\n", hsv(0.5, 1, 1))'

 hsv(hue=0.5, saturation=1, value=1)

Additionally you might binarize with Otsu, paying attention just to the bluishness of pixels rather than the grey-scale intensity of pixels. The Open CV library has a good Otsu implementation. And then you can evaluate your masked image, or perhaps simply examine the resulting mask.