TopolearnMost algorithms want grayscale, other libraries want RGB, and colour segmentation works far better in HSV.
cv2.cvtColor(img, code) converts an image between colour spaces. The codes read like COLOR_<from>2<to>:
cv2.COLOR_BGR2GRAY: one channel, a weighted sum 0.299R + 0.587G + 0.114Bcv2.COLOR_BGR2RGB: needed before handing images to matplotlib, Pillow or most deep learning modelscv2.COLOR_BGR2HSV: hue, saturation, valuecv2.COLOR_BGR2LAB: perceptually more uniform; good for colour distances and contrast tweaks
HSV separates which colour (hue) from how strong (saturation) and how bright (value). A red ball under shade and in sunlight has very different BGR values but a similar hue, which is why colour thresholding is done in HSV.
For 8-bit images OpenCV squeezes hue into 0 to 179 (degrees divided by 2) so it fits in a byte; S and V use 0 to 255. Red sits at both ends of the hue circle, around 0 and around 179, so selecting red needs two ranges.
Gray to BGR (COLOR_GRAY2BGR) does not invent colour; it copies the one channel three times so you can draw coloured annotations on it.
Flipping channel order with slicing, img[..., ::-1], also converts BGR to RGB, but it returns a non-contiguous view that some cv2 functions reject; cvtColor always returns a fresh contiguous array.
Example
import cv2
import numpy as np
img = np.zeros((1, 3, 3), np.uint8)
img[0] = [(0, 0, 255), (0, 255, 0), (255, 0, 0)] # red, green, blue (BGR)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
print(gray) # [[ 76 150 29]]
rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
print(rgb[0, 0]) # [255 0 0] red is first now
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
print(hsv[0, :, 0]) # [ 0 60 120] hue in 0..179 (degrees / 2)
print(hsv[0, :, 1:]) # saturation and value are 255Remember
cv2.cvtColor(img, cv2.COLOR_X2Y)converts colour spaces- Convert BGR to RGB before passing images to other libraries
- HSV separates colour from brightness; 8-bit hue runs 0 to 179
- Red wraps around hue 0/179, so it needs two ranges
Watch out for
- Using hue thresholds in degrees (0 to 360) on 8-bit HSV images, where hue is halved.