对齐人脸图像并在 python cv2 中合并

Question

我有一堆人脸图像数据集（取自 http://vision.ucsd.edu /content/yale-face-database ），我基本上想把它从电影扫描仪中转变成黑暗的打乱套装的 gif（ http://2.bp.blogspot.com/-tRLWSOqh84Y/VSb_cF7sOoI/AAAAAAAAAWI/3XqT6d_exso/s1600/scramble%2Bsuit%2B2.gif）。 到目前为止，我能够拍摄图像并用 python 将它们批量切割成面部“碎片”。 我无法做的下一步是“对齐”这些面，以便所有部分在合并或放回到一起时形成一个面。 我也不确定如何将它们合并或重新组合在一起。 一旦我有了一堆随机拼凑在一起的图像，我就可以自己创建 gif 了。

这是我到目前为止拍摄图像、将它们转换为 jpg 并将它们切割成必要的部分的代码（取自此处 https://leslietj.github.io/2020/06/30/Automatic-Face-Crop-Using-Dlib/ ）：

import sys
import dlib
from skimage import io
import numpy as np
import cv2
import matplotlib.pylab as plt
import math
from PIL import Image
import os

def arc_points(point1, point2, num_of_points):
    points = []
    center_x = (point1[0] + point2[0])/2
    center_y = (point1[1] + point2[1])/2
    radius = abs((point1[0] - point2[0])/2)
    for i in range(num_of_points):
        if i == 0:
            continue
    
        point = []
        x = center_x + radius * math.cos(math.pi + i * math.pi / num_of_points)
        y = center_y + radius * math.sin(math.pi + i * math.pi / num_of_points)
        point.append(x)
        point.append(y)
        
        points.append(point)
    
    return points


def get_landmarks(img,mode=1):
    dets = detector(img, 1)
    landmarks = np.zeros((34, 2))
    for k, d in enumerate(dets):
        shape = predictor(img, d)

        #quarter face (#1)
        if mode == 1:
            landmarks[0]= (shape.part(0).x, shape.part(0).y)
            landmarks[1] = (shape.part(1).x, shape.part(1).y)
            landmarks[2] = (shape.part(2).x, shape.part(2).y)
            landmarks[3] = (shape.part(30).x, shape.part(30).y)
            landmarks[4] = (shape.part(29).x, shape.part(29).y)
            landmarks[5] = (shape.part(28).x, shape.part(28).y)
            point1 = [shape.part(0).x, shape.part(0).y]
            point2 = [shape.part(28).x, shape.part(28).y]
            points = arc_points(point1, point2, 29)
            for i in range(len(points)):
                landmarks[33 - i] = (points[i][0], points[i][1])
        

        #half face (#2)

        if mode == 2:
            landmarks[0] = (shape.part(0).x, shape.part(0).y)
            landmarks[1] = (shape.part(1).x, shape.part(1).y)
            landmarks[2] = (shape.part(2).x, shape.part(2).y)
            landmarks[3] = (shape.part(14).x, shape.part(14).y)
            landmarks[4] = (shape.part(15).x, shape.part(15).y)
            landmarks[5] = (shape.part(16).x, shape.part(16).y)
            point1 = [shape.part(0).x, shape.part(0).y]
            point2 = [shape.part(16).x, shape.part(16).y]
            points = arc_points(point1, point2, 29)
            #print(points)
            for i in range(len(points)):
                #print(33-i)
                landmarks[33 - i] = (points[i][0], points[i][1])
        
        if mode == 3:
            #3/4 face (#3)
            for i in range(9):
                landmarks[i] = (shape.part(i).x, shape.part(i).y)
            landmarks[9] = (shape.part(31).x, shape.part(31).y)
            
            landmarks[10] = (shape.part(14).x, shape.part(14).y)
            landmarks[11] = (shape.part(15).x, shape.part(15).y)
            landmarks[12] = (shape.part(16).x, shape.part(16).y)
            point1 = [shape.part(0).x, shape.part(0).y]
            point2 = [shape.part(16).x, shape.part(16).y]
            points = arc_points(point1, point2, 22)
            for i in range(len(points)):
                landmarks[33 - i] = (points[i][0], points[i][1])
        #full face (#4)
        if mode == 4:
            for i in range(17):
                landmarks[i] = (shape.part(i).x, shape.part(i).y)
            point1 = [shape.part(0).x, shape.part(0).y]
            point2 = [shape.part(16).x, shape.part(16).y]
            points = arc_points(point1, point2, 18)
            for i in range(len(points)):
                landmarks[33 - i] = (points[i][0], points[i][1])
        

    return landmarks


def inside(X,Y,Region): 
    j=len(Region)-1
    flag=False
    for i in range(len(Region)):
        if (Region[i][1]<Y and Region[j][1]>=Y or Region[j][1]<Y and Region[i][1]>=Y):  
            if (Region[i][0] + (Y - Region[i][1]) / (Region[j][1] - Region[i][1]) * (Region[j][0] - Region[i][0]) < X):
                flag =not flag
        j=i
    return flag

count=0
files = os.listdir('yalefaces')
for filename in files:
    if filename.endswith('glasses') or filename.endswith('happy') or filename.endswith('noglasses') or filename.endswith('normal'):
        path = os.path.join('yalefaces',filename)

  
        # importing the image 
        im = Image.open(path)

      
        # converting to jpg
        rgb_im = im.convert("RGB")
      
        # exporting the image
        rgb_im.save('temp.jpg')

        count+=1


        path = 'temp.jpg'
        for im in range(1,5):
            #path = 'subject01.jpg'
            detector = dlib.get_frontal_face_detector() 
        # the .dat file can be downloaded following this link:
        # https://sourceforge.net/projects/dclib/files/dlib/v18.10/shape_predictor_68_face_landmarks.dat.bz2/download
            predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
            img = io.imread(path)

            region = get_landmarks(img,mode=im)
            shape = list(img.shape) 
            cropped_img = img.copy()
            for i in range(shape[0]):
                for j in range(shape[1]):
                    if not inside(j, i, region):
                        #print(img[0])
                        cropped_img[i, j] = (img[0,0][0], img[0,0][1], img[0,0][2]) # the RGB values of the background


            cropped_img = cv2.cvtColor(cropped_img, cv2.COLOR_BGR2GRAY)
            cv2.imwrite(str(count).zfill(3)+'-'+str(im).zfill(2)+'.jpg', cropped_img)

Answer 1

因为这是预处理的几个步骤：

1. 使用此脚本对齐面部 https://pyimagesearch.com/2017/05/22/face-alignment-with-opencv-and-python/
2. 将面部切碎并将它们放在一起。如前所述，我正在做 cv2.add 这不是我想要的。我实际上希望各层堆叠起来，如果下面有任何东西，请忽略它。 cv2.add 不这样做，所以我必须自己制作。因此，如果我有两层，并且我想将它们堆叠起来，以便顶部的第一层优先，并且如果上面有任何内容，则忽略底层。

def reduction(layer1,layer2):
    for i in range(0,layer1.shape[0]):
        for j in range(0,layer1.shape[1]):
            pixel1 = layer1.item(i, j)
            pixel2 = layer2.item(i, j)
            if layer2[i,j] != 255:
                layer1[i,j]=255
    return layer1

layer2 = reduction(layer2,layer1)

for i in range(0,layer1.shape[0]):
    for j in range(0,layer1.shape[1]):
        pixel = layer2[i,j]
        if pixel != 255:
            layer1[i,j]=layer2[i,j]

就是这样。我没有意识到图像只是 numpy 数组，所以我可以直接操作数组。