dlib实现红眼特效

效果

（图片来源网络，如有侵权请联系删除）

在这里插入图片描述

摘要

技术栈如下：

代码演示：python
面部关键点识别：dlib
图像处理：pillow，opencv

人脸关键点

既然是做人眼部位的贴图，需要知道人眼与水平方向的夹角，关键点就是左眼的36点和39点，右眼的42点和45点。

在这里插入图片描述
还需要准备红眼贴纸，用于后续覆盖在眼部。

代码

import cv2 as cv
import dlib
from PIL import Imagedef detect_predict():# 生成predictor和detectordetector = dlib.get_frontal_face_detector()predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')return detector, predictordef cal_ang(point_1, point_2, point_3):# 三点计算夹角，返回角度对应各点的角度a = math.sqrt((point_2[0]-point_3[0])*(point_2[0]-point_3[0])+(point_2[1]-point_3[1])*(point_2[1] - point_3[1]))b = math.sqrt((point_1[0]-point_3[0])*(point_1[0]-point_3[0])+(point_1[1]-point_3[1])*(point_1[1] - point_3[1]))c = math.sqrt((point_1[0]-point_2[0])*(point_1[0]-point_2[0])+(point_1[1]-point_2[1])*(point_1[1]-point_2[1]))A = math.degrees(math.acos((a*a-b*b-c*c)/(-2*b*c)))B = math.degrees(math.acos((b*b-a*a-c*c)/(-2*a*c)))C = math.degrees(math.acos((c*c-a*a-b*b)/(-2*a*b)))return math.ceil(A), math.ceil(B), math.ceil(C)origin = cv.imread('人脸图片')
gray = cv.cvtColor(origin, cv.COLOR_BGR2GRAY)detector, predictor = detect_predict()
res = detector(gray, 1)faces = []
for face in res:shape = predictor(origin, face)parts = shape.parts()faces.append(parts)scale = 5.2      # mask宽度 = 眼距 * scale
pil_img = Image.fromarray(cv.cvtColor(origin, cv.COLOR_BGR2RGB))
mask = Image.open('红眼贴纸')
mask = mask.convert('RGBA')
mw, mh = mask.sizemeasure_l = 0.55       # 左眼mask的中心点位置
measure_r = 0.48       # 右眼mask的中心点位置
eyes = []# 左眼
angle = cal_ang((p[36].x, p[36].y), (p[36].x + 10, p[36].y), (p[39].x, p[39].y))
ang = int('-' + str(angle[0])) if p[36].y < p[39].y else angle[0]
canthus = math.sqrt(math.pow(p[36].x - p[39].x, 2) + math.pow(p[36].y - p[39].y, 2))
target_h = int(mh / (mw / (canthus * scale)))
m1 = mask.resize((int(canthus * scale), target_h), Image.ANTIALIAS)
m1 = m1.rotate(ang, expand=True)
w, h = m1.size
x, y = ((p[38].x + p[40].x) / 2, (p[38].y + p[40].y) / 2)# 小于0时，会出现1px的偏差
ltx = int(x - measure_l * w) if int(x - measure_l * w) >= 0 else int(x - measure_l * w) - 1
lty = int(y - h/2 + ang / 180 * canthus) if int(y - h/2 + ang / 180 * canthus) >= 0 else int(y - h/2 + ang / 180 * canthus) - 1
position = (ltx, lty, int(x - w * measure_l + w), int(y - h/2 + h + ang / 180 * canthus))
pil_img.paste(m1, position, mask=m1.split()[-1])# 右眼
angle = cal_ang((p[42].x, p[42].y), (p[42].x + 10, p[42].y), (p[45].x, p[45].y))
ang = int('-' + str(angle[0])) if p[42].y < p[45].y else angle[0]
canthus = math.sqrt(math.pow(p[42].x - p[45].x, 2) + math.pow(p[42].y - p[45].y, 2))
target_h = int(mh / (mw / (canthus * scale)))
m2 = mask.resize((int(canthus * scale), target_h), Image.ANTIALIAS)
m2 = m2.rotate(ang, expand=True)
w, h = m2.size
x, y = ((p[43].x + p[47].x) / 2, (p[43].y + p[47].y) / 2)# 小于0时，出现1px的偏差
ltx = int(x - measure_r * w) if int(x - measure_r * w) >= 0 else int(x - measure_r * w) - 1
lty = int(y - h / 2 - ang / 180 * canthus) if int(y - h / 2 - ang / 180 * canthus) >= 0 else int(y - h / 2 - ang / 180 * canthus) - 1
position = (ltx, lty, int(x - w * measure_r + w), int(y - h / 2 + h - ang / 180 * canthus))
pil_img.paste(m2, position, mask=m2.split()[-1])