List of articles

• One 、 Principle of face illumination transfer
• Two 、python Realize face illumination migration

Recently, the main work is to evaluate the quality of human face illumination . I originally wanted to use the traditional block based + In the form of variance , This method has some limitations , Can't adapt to all situations . then ,github Find some related projects , Using the method of deep learning .

FIIQA-PyTorch

FIIQA

Be happy , Run , I found it completely useless . The data downloaded is only 3 class , combination github On 200 class , There is a little gap .

So I want to make it myself , The reference papers are as follows ：

Face Illumination Transfer through Edge-preserving Filters

One 、 Principle of face illumination transfer

In this paper, faces are mainly divided into 3 part ： Color layer 、 Large scale layer and detail layer . Convert the face image to Lab Color space , among L The channel represents the brightness value ;a、b Represents the color layer . Yes L Passage through wsl（ Weighted least two filter ） Large scale layer can be obtained by filtering , And then through the following ：
d e t a i l s = L l a r g e s c a l e details=\frac{L}{largescale} details=largescaleL​
among ：largscale Represents a large-scale layer ,details Represents the detail layer .

By the above methods , Extract large-scale layers from the template and the target image in turn , then , These two images are Guided filtering / Guided filtering （Guided Filter）. Be careful , Here, take the target as the guidance diagram , Template as initial image , Make the template similar to the target .

adopt , After guidance filtering , You need to restore the brightness channel L, Calculate with reference to the above formula , And then Lab Space is transformed into BGR Space , Finally, we get the migration picture of uneven illumination of human face .

Two 、python Realize face illumination migration

import cv2
from cv2.ximgproc import *
import numpy as np
def lightness_layer_decomposition(img,conf,sigma):
''' Separate the picture into color layer and brightness layer :param img: :return: '''
# wsl wave filtering 
large_scale_img=fastGlobalSmootherFilter(img, img, conf, sigma)
detail_img=img/large_scale_img
return large_scale_img,detail_img
def face_illumination_transfer(target=None, reference=None):
''' Migrate the label illumination to the target face :param target: :param reference: :return: '''
h,w=reference.shape[:2]
target=cv2.resize(target,(w,h))
# lab Color conversion 
# Extract color (a,b) And brightness layer l
lab_img = cv2.cvtColor(target, cv2.COLOR_BGR2Lab)
l, a, b = cv2.split(lab_img)
lab_rimg = cv2.cvtColor(reference, cv2.COLOR_BGR2Lab)
lr, ar,br = cv2.split(lab_rimg)
# Layer the brightness layer : Large scale layer and detail layer 
large_scale_img, detail_img= lightness_layer_decomposition(l,600,20)
large_scale_rimg, detail_rimg = lightness_layer_decomposition(lr,600,20)
# The template brightness is transferred to the target image by guided wave filtering 
large_scale_rimg = large_scale_rimg.astype('float32')
large_scale_img = large_scale_img.astype('float32')
out=guidedFilter(large_scale_img,large_scale_rimg,18,1e-3)
out=out*detail_img
out=out.astype(np.uint8)
res=cv2.merge((out,a,b))
res=cv2.cvtColor(res,cv2.COLOR_Lab2BGR)
return res
if __name__=="__main__":
img_file=r"D:\data\face\good3\1630905386503.jpeg"
reference=r"D:\data\face_illumination\illumination patterns\office-lamp-night-6.JPG"
# Read face pictures 
img = cv2.imread(img_file)
rimg=cv2.imread(reference)
res=face_illumination_transfer(img,rimg)
cv2.imshow("res",res)
cv2.waitKey(0)

The final results are as follows ：

Refer to the connection ：

opencv Use — fastGlobalSmootherFilter

Pilot map filtering （Guided Image Filtering） Principles and OpenCV Realization