9

I have seen all questions on SO for range of HSV color space for skin
But I can only figure out this

Code -

CvScalar  hsv_min = cvScalar(0, 30, 60, 0);
CvScalar  hsv_max = cvScalar(20, 150, 255, 0);
//range I am using is { 0,30,60,0 & 20,150,255,0 }
cvCvtColor(src, hsv_image, CV_BGR2HSV);
cvInRangeS (hsv_image, hsv_min, hsv_max, hsv_mask);
cvDilate(hsv_mask,hsv_mask,0,1);
cvErode(hsv_mask,hsv_mask,0,1);
cvSmooth( hsv_mask, hsv_mask, CV_MEDIAN);

Problem with this range ( { 0,30,60,0 & 20,150,255,0 } ) is it detects even red color and when you place your hand in red background it does not track your skin...
Please Help !!!

Oli
  • 15,935
  • 7
  • 50
  • 66
Wazy
  • 8,822
  • 10
  • 53
  • 98
  • 1
    What about African Americans (if that's even the right term any more)? Or native Americans? Or Indians? Or Australian Aborigines? Or little green men? – paxdiablo Jan 06 '12 at 05:41
  • I am talking general...It should be for human skin...Please give me different ranges if that's the case...I think we can use cvInRangeS for multiple ranges... – Wazy Jan 06 '12 at 05:46
  • Therein lies your problem. Since you're only interested in humans (and I assume you don't subscribe to some bizarre racist philosophy of considering other races subhuman), we can only count out the little green men. That still leaves a _massive_ color range to cater for. Now I'm assuming this is to do something with photos or videos (telling the difference between skin and everything else) so, unless your photo is tagged somehow as containing only caucasions or Aborigines or whatnot, the chances of being able to tell the difference between human and non-human is small. Even with multi-ranges. – paxdiablo Jan 06 '12 at 05:51
  • 1
    Now this _may_ be doable if there are more constraints or I have the assumption of your actual need quite wrong. In which case, let us know. – paxdiablo Jan 06 '12 at 05:52
  • @paxdiablo I want to track hand...Basically AR of watches...In that user will keep his hand in a template image of hand and we get the distance of wrist from figures. What we can do more is let user click at specific position and get hsv values...Please help me.... – Wazy Jan 06 '12 at 05:57
  • Please Help me with the range.... – Wazy Jan 06 '12 at 11:41

9 Answers9

9

According to this : http://matmidia.org/sibgrapi2009/media/posters/59928.pdf

The skin in channel H is characterized by values between 0 and 50, in the channel S from 0.23 to 0.68 for Asian and Caucasian ethnics.

You'll have to take extra care with very dark parts of the image and probably discard them altogether, as the HSV conversion gets really noisy for small values of V.

Depending on what your constraints are, you could also consider using a coloured glove (some colour that is not appearing normally in the scene), or setting a background of different colour than red, that is further away from skin colours (magenta, green, whatever).

Leonidas K
  • 491
  • 2
  • 8
  • 2
    Link is dead, but [this](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.573.9432&rep=rep1&type=pdf) seems to be the same article. – Junuxx May 06 '21 at 20:28
7

I have tried
lower = np.array([0, 10, 60], dtype = "uint8") upper = np.array([20, 150, 255], dtype = "uint8") It gives almost good result .

Mayank Kumar
  • 523
  • 6
  • 9
7

Basically, it's hard to have one fixed color range for skin, because even if you want to detect only your own skin, its color will actually change a lot depending on lighting conditions.

So, maybe you can use the idea of this nice scientific article from 2011:

http://www.robots.ox.ac.uk/~vgg/research/hands/

Basically, they detect face (it's easy with oepncv). Then they extract the skin color of the face (which is very specific to the persons on the image). Then they detect skin using this color. Since the color is very specific, they should have much less false detection, than what you have with your fixed color range.

Oli
  • 15,935
  • 7
  • 50
  • 66
  • Just read your comment and liked the idea ...coz that what I am trying to do...Let me check out the link you provided...Thanks a lot for the answer.... – Wazy Jan 06 '12 at 11:34
5

I know it is too late to reply to this. But I am doing the same, and I used K means clustering to get the skin color. That is first you have to Detect the face, which I am doing with the use of Haar cascade classifier, and then based on the coordinates of the face you can crop the face and then use it as the source to cluster the color. Find out the cluster that has the most elements and that will be your skin color. Or without specifying one particular value you can get the cluster center point and you can use it as a range by deducting and adding a particular value. This will be helpful http://answers.opencv.org/question/23196/k-mean-clustering-of-hsv-histogram-of-frames-of-a-video/

nickY
  • 95
  • 3
  • 9
3

Try this preprocessing function in your code:

def preprocess(action_frame):

    blur = cv2.GaussianBlur(action_frame, (3,3), 0)
    hsv = cv2.cvtColor(blur, cv2.COLOR_RGB2HSV)

    lower_color = np.array([108, 23, 82])
    upper_color = np.array([179, 255, 255])

    mask = cv2.inRange(hsv, lower_color, upper_color)
    blur = cv2.medianBlur(mask, 5)

    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (8, 8))
    hsv_d = cv2.dilate(blur, kernel)

return hsv_d

I have been using this for months now. It might solve your problem forever.

Ashish Gupta
  • 185
  • 1
  • 12
3

Adaptive skin detector exist on OpenCv see samples/c/adaptiveskindetector_sample.cpp

Birol Kuyumcu
  • 1,195
  • 7
  • 17
1

And if you are using video caption, you should be concerned that in most of the cases the cameras make an automatic contrast adjust, this impact in your images because you are going to have a significant difference of tones, also you should consider the shadows and light sources that impact in the skin

Alexander Leon VI
  • 499
  • 1
  • 4
  • 18
1

Deepgaze is a library with skin detecting in addition to other functionalities.

DeepGaze Github

He uses:

[0, 58, 50] lower bound skin HSV
[30, 255, 255] upper bound skin HSV

He is quite responsive on Github so an option would be to contact him and ask how he came up with those numbers.

mLstudent33
  • 1,033
  • 3
  • 14
  • 32
1

This is one of the ideal solution for skin detection..

Most of the answers above work for variety of skin colors like dark red, light yellow, light orange, .... But these colors are not supposed to be skin colors as they are barely seen in images.

As you can see in this pic, the red range is somewhere between 350 to 20. enter image description here

So there are two possible ranges of Hue for skins.

HSV range

#1st range of Hue
lower -> [0, 30, 53]
upper -> [20, 150, 255]

#2nd range of Hue (OpenCV converts 360 to 180)
lower2 -> [172, 30, 53]
upper2 -> [180, 180, 210]

Full code:

import os, cv2
import numpy as np

image = cv2.imread(os.path.join('skin.png'))

# Covers both range
lower = np.array([0, 30, 53], dtype = "uint8")
upper = np.array([20, 180, 255], dtype = "uint8")

lower2 = np.array([172, 30, 53], dtype = "uint8")
upper2 = np.array([180, 180, 210], dtype = "uint8")

converted = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)

skinMask = cv2.inRange(converted, lower, upper)
skinMask2 = cv2.inRange(converted, lower2, upper2)

#Gaussian Blur
skinMask = cv2.GaussianBlur(skinMask, (3, 3), 0)
skinMask2 = cv2.GaussianBlur(skinMask2, (3, 3), 0)

skin1 = cv2.bitwise_and(image, image, mask = skinMask)
skin2 = cv2.bitwise_and(image, image, mask = skinMask2)
skin = cv2.bitwise_or(skin1,skin2) #adding both ranges

# show the skin in the image along with the mask
cv2.imshow("images", np.hstack([frame, skin]))
cv2.waitKey(0)
Prakash Dahal
  • 4,388
  • 2
  • 11
  • 25