前言

人脸识别出五官后，可以增加一些形变的算法，对人脸进行“整形”，比如利用凸面镜的滤镜效果，来使得眼睛增大，起到美颜或者夸张艺术效果，效果如下（下面有点夸张了勿喷，可适当调节滤镜范围）：

原理

整个流程是，利用人脸识别算法库（如face_recognition），识别出人脸的两只眼睛的中心坐标，并利用类似局部“凸透镜效果算法”,即：（相当于图像一片区域通过凸透镜所看到的视觉效果。 本质是一种插值算法，就是沿着图像区域的中心（透镜中心）将像素径向放大，离中心近的物体得权重大，并且向边缘处挤压（边缘近的像素被压缩），同事在边缘处需要连续，这样图像看起来会自然些），从眼睛中心位置进行逐个像素变形，这里我使用了椭圆型的滤镜（直觉上感觉眼睛接近椭圆用椭圆比较自然点），区别于网上的圆形滤镜，核心代码如下：

def bigeye(src_img,eye,r=20,R=30):   H=src_img.shape[0]#yW=src_img.shape[1]#xnew_img=src_img.copy()#print(new_img.shape)              center_x=eye[0]center_y=eye[1]for i in range(W):for j in range(H):distance=((i-center_x)*(i-center_x)+(j-center_y)*(j-center_y))new_dist=math.sqrt(distance)if (i-center_x)**2/(R**2)+(j-center_y)**2/(r**2)<1:u=int(np.floor(new_dist*(i-center_x)/syd(center_x,center_y,R,r,i,j)+center_x))v=int(np.floor(new_dist*(j-center_y)/syd(center_x,center_y,R,r,i,j)+center_y))#u=int(int(i-center_x)/2*(new_dist/r)+center_x)#v=int(int(i-center_y)/2*(new_dist/r)+center_y)if 0<u<W and 0<v<H:new_img[j,i,:]=src_img[v,u,:]              new_img=sydd(new_img,int(W/2),int(H/2),R,r)return new_img

同时利用opencv的ui建两个滑块，用于调节椭圆参数，这样就可以通过调节椭圆的大小半径对眼睛区域进行必要的“椭圆放大”。可以通过调节找到合适的参数，在实际的应用中就不会向本篇图片那么夸张了。
完整的代码如下：

Python源码实现

# -*- coding: utf-8 -*-
"""
Created on Tue Apr  5 22:19:08 2022@author: JAMES FEI 
Copyright (C) 2022 FEI PANFENG, All rights reserved.
THIS SOFTEWARE, INCLUDING DOCUMENTATION,IS PROTECTED BY COPYRIGHT CONTROLLED 
BY FEI PANFENG ALL RIGHTS ARE RESERVED."""
import face_recognition
import cv2
import numpy as np
import mathvideo_capture = cv2.VideoCapture(0)#img=cv2.imread('55.JPG')def syd(cx,cy,a,b,i,j):"""计算椭圆半径"""p=np.arctan2((i-cx),-(j-cy))if b<=0:b=1t=np.arctan(np.tan(p)*a/b)x=a*np.cos(t)y=b*np.sin(t)r=np.sqrt(x**2+y**2)return rdef sydd(img,cx,cy,a,b):"""计算椭圆半径"""for i in range(180):        t=2*np.pi/359x=int(a*np.cos(t*i)+cx)y=int(-b*np.sin(t*i)+cy)cv2.circle(img, (x,y), 2, (0,255,0), 1)          return imgdef bigeye(src_img,eye,r=20,R=30):   H=src_img.shape[0]#yW=src_img.shape[1]#xnew_img=src_img.copy()#print(new_img.shape)              center_x=eye[0]center_y=eye[1]for i in range(W):for j in range(H):distance=((i-center_x)*(i-center_x)+(j-center_y)*(j-center_y))new_dist=math.sqrt(distance)if (i-center_x)**2/(R**2)+(j-center_y)**2/(r**2)<1:u=int(np.floor(new_dist*(i-center_x)/syd(center_x,center_y,R,r,i,j)+center_x))v=int(np.floor(new_dist*(j-center_y)/syd(center_x,center_y,R,r,i,j)+center_y))#u=int(int(i-center_x)/2*(new_dist/r)+center_x)#v=int(int(i-center_y)/2*(new_dist/r)+center_y)if 0<u<W and 0<v<H:new_img[j,i,:]=src_img[v,u,:]              new_img=sydd(new_img,int(W/2),int(H/2),R,r)return new_img
def nothing(x):pass
cv2.namedWindow('bigeye')
cv2.createTrackbar('R', 'bigeye', 10, 100, nothing)
cv2.createTrackbar('r', 'bigeye', 5, 100, nothing)while True:r = cv2.getTrackbarPos('r', 'bigeye')R = cv2.getTrackbarPos('R', 'bigeye')if r<=0:r=1if R<=0:R=1#frame=img.copy()ret, frame = video_capture.read()#small_frame = cv2.resize(frame, (0, 0), fx=1/4, fy=1/4)rgb_frame = frame[:, :, ::-1]face_landmarks_list = face_recognition.face_landmarks(rgb_frame)output=framefor face_landmarks in face_landmarks_list:lef_center_x=0lef_center_y=0        for point in face_landmarks["left_eye"]:lef_center_x+=point[0]lef_center_y+=point[1]lef_center_x=int(lef_center_x/len(face_landmarks["left_eye"]))lef_center_y=int(lef_center_y/len(face_landmarks["left_eye"]))#cv2.circle(frame, (lef_center_x,lef_center_y), 30, (0,0,0), -1)  if lef_center_x and lef_center_y:output=bigeye(output,(lef_center_x,lef_center_y),r=r,R=R)  rig_center_x=0rig_center_y=0for point in face_landmarks["right_eye"]:rig_center_x+=point[0]rig_center_y+=point[1]rig_center_x=int(rig_center_x/len(face_landmarks["left_eye"]))rig_center_y=int(rig_center_y/len(face_landmarks["left_eye"]))#cv2.circle(frame, (rig_center_x,rig_center_y), 30, (0,0,0), -1)  if rig_center_x and rig_center_y:output=bigeye(output,(rig_center_x,rig_center_y),r=r,R=R)  cv2.imshow('bigeye', output)# 按Q退出if cv2.waitKey(1) & 0xFF == ord('q'):breakvideo_capture.release()
cv2.destroyAllWindows()