Calculate color visibility from hue values respecting saturation and value in image numpy array

Question

For a fun project I want to analyze a few images, especially which colors (hue) are more visible than others. As I want to take the “visibility” of the colors in account, just counting the hue of pixels is not enough (e.g. perfect black would count as red as its hue is 0°). I came up with a formula which is IMO good enough for my project.

Currently I do the following:

Read the image with opencv (results in BGR numpy array)
Translate the image to HSV
For each pixel, calculate the visibility of its hue (from saturation and value) and sum it in a dict of hues.

The formula is color_visibility = sqrt(saturation * value). So a full-red RGB=255,0,0; HSV=0,1,1 would result in 1 while e.g. a light-red RGB=255,128,128; HSV=0,0.5,1 would result in 0.70.

Here is the (full working) code I use:

import urllib
import cv2
import numpy as np

url = 'https://upload.wikimedia.org/wikipedia/commons/thumb/0/02/Leuchtturm_in_Westerheversand_crop.jpg/299px-Leuchtturm_in_Westerheversand_crop.jpg'
image = np.asarray(bytearray(urllib.urlopen(url).read()), dtype="uint8")
image = cv2.imdecode(image, cv2.IMREAD_COLOR)

d = {}
hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
pixels = hsv.reshape((hsv.shape[0] * hsv.shape[1], 3))
for h,s,v in pixels:
    d[h] = d.get(h, 0.) + (s/255. * v/255.) ** 0.5

As you might guess, the code gets really slow when the image have more pixels.

My question is, how can I do the calculation of my formula without the dict and for-loop? Maybe directly with numpy?

Asked By: sphere

||

Source

Answer 1

The magic you are looking for is in np.bincount, as it translates pretty straight-forwardly to the loopy version using the h values as the bins –

H,S,V = pixels.T
d_arr = np.bincount(H, ((S/255.0) * (V/255.0))**0.5 )

Note that the resultant array might have elements have zero valued counts

Answered By: Divakar

Answer 2

This solution is almost what I need, but what I wanted to return was the total percentage of pixels that exceed a BASELINE value (of say, .80). I was trying to do this when a lambda on the resulting d_arr, but just got a mess back. Any thoughts on how to just return the single value for the entire image?

pixels = hsv.reshape((hsv.shape[0] * hsv.shape[1], 3))
    for h,s,v in pixels:
      d[h] = np.sqrt(s/255. * v/255.)
      cooked_px[h] = filter(lambda x: x > BASELINE, d)

Answered By: sed

Calculate color visibility from hue values respecting saturation and value in image numpy array

Question:

Answers: