Improve face-detection based skin tone filter (Python OpenCV)

Question

I'm using Python OpenCV to implement an adaptive skin color filter that uses haarcascades to detect an upright face, followed by filtering the face ROI to remove non-skin features like eyebrows, glasses etc to get the average skin tone (in RGB). Then I convert the image to HSV and extract the HSV values close to the average I obtained. Here is my code:

import cv2
import numpy as np
from functions import *
def nothing(x):
    pass
cap = cv2.VideoCapture(0)
face_cascade = cv2.CascadeClassifier('haarcascade_frontalface_alt.xml')
cv2.namedWindow('Video')
cv2.moveWindow('Video',5,5)
cv2.namedWindow('HSV_Thresh')
cv2.moveWindow('HSV_Thresh',655,5)
cv2.createTrackbar('tval', 'Video', 29, 255, nothing)
cv2.createTrackbar('htoler', 'HSV_Thresh', 17, 100, nothing)
cv2.createTrackbar('stoler', 'HSV_Thresh', 25, 100, nothing)
cv2.createTrackbar('vtoler', 'HSV_Thresh', 84, 100, nothing)

kernel = np.ones((5, 5), np.uint8)# 5X5 erosion kernel
bavg=0
ravg=0
gavg=0
while True: 
    tval1=cv2.getTrackbarPos('tval', 'Video')#thresh value to remove non skin components from face
    htoler_val=cv2.getTrackbarPos('htoler', 'HSV_Thresh')
    stoler_val=cv2.getTrackbarPos('stoler', 'HSV_Thresh')
    vtoler_val=cv2.getTrackbarPos('vtoler', 'HSV_Thresh')
    ret,img=cap.read()#Read from source
    img[0:100,0:100] = [255,255,255]
    thresh_hsv_toler=img    
    faces = face_cascade.detectMultiScale(img, 1.3, 5)
        for (x,y,w,h) in faces:
        bavg=0
        ravg=0
        gavg=0
        numpix=0
            cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
        cv2.rectangle(img,(x,y),(x+w,y+h),(0,255,0),2)
        roi_face = img[y:y+h, x:x+w]
        #avg_col=img[100,100]
        rect_face=img[y:y+h-h/8,x+w/7:x+w-w/5]#extract only skin features from remaining bg
        mask=cv2.inRange(rect_face,(tval1,tval1,tval1),(255,255,255))
        mask=cv2.cvtColor(mask,cv2.COLOR_GRAY2BGR)
        tone=cv2.subtract(mask,rect_face)
        tone=cv2.subtract(mask,tone)
        (rows,cols,col)=tone.shape # 480 rows and 640 cols; 3 values for RGB img
        for i in range(rows): #note the presence of colon
            for j in range(cols):
                if (tone[i,j,0]!=0 and tone[i,j,0]!=0 and tone[i,j,0]!=0):
                    bavg=bavg+tone[i,j,0]
                    gavg=gavg+tone[i,j,1]
                    ravg=ravg+tone[i,j,2]
                    numpix=numpix+1
        bavg=bavg/numpix
        gavg=gavg/numpix
        ravg=ravg/numpix
        '''print "bavg="+str(bavg)
        print "gavg="+str(gavg)
        print "ravg="+str(ravg)
        print "numpix="+str(numpix)'''
        cv2.circle(img, (50,50), 20, (bavg,gavg,ravg), 50)#get obtained average colour on screen


        cv2.imshow('skin_mask', tone)
        hsv=cv2.cvtColor(img,cv2.COLOR_BGR2HSV)
        thresh_hsv_toler=cv2.inRange(hsv,(hsv[50,50,0]-htoler_val,hsv[50,50,1]-stoler_val,hsv[50,50,2]-vtoler_val),(hsv[50,50,0]+htoler_val,hsv[50,50,1]+stoler_val,hsv[50,50,2]+vtoler_val))

        thresh_hsv_toler=cv2.dilate(thresh_hsv_toler, kernel, iterations=1)
        thresh_hsv_toler=cv2.cvtColor(thresh_hsv_toler,cv2.COLOR_GRAY2BGR)#superimposing binary mask on image
        hsv_filter=cv2.subtract(thresh_hsv_toler,img)
        hsv_filter=cv2.subtract(thresh_hsv_toler,hsv_filter)



        cv2.imshow('HSV_Thresh', hsv_filter)


    if(cv2.waitKey(10) & 0xFF == ord('b')):
            break
    cv2.imshow('Video', img)

Here is the output:

How do I reduce the noise in my result without adding a second filter ? Thanks

Answer 1

your tolerance value for BG and R while extracting the face mask are uniform. You might need to include separate tolerance thresholds for the colour spaces, eg. more tolerance value for red