手把手教你搭建明星脸相似度分析系统

之前和女朋友在微信上玩明星脸小程序，发现大多小程序的分析都不太准，偶尔有几个准的还收费，正好之前学过人脸识别，想着原理应该大同小异，就决定自己搭建一个明星脸程序。
github项目地址：https://github.com/JiageWang/starface

1. 数据收集

要寻找最相似明星脸，首先得有数据，因为现成的数据集大多过时了，缺少很多当红的明星，因此决定自己去网上爬取，找了很多网站最终决定对明星网下手，写了个简单的脚本如下，主要使用的requests与xpath技术

import os
import time
import random
import requests
from lxml import etreefirst_num = random.randint(55, 62)
third_num = random.randint(0, 3200)
fourth_num = random.randint(0, 140)os_type = ['(Windows NT 6.1; WOW64)', '(Windows NT 10.0; WOW64)', '(X11; Linux x86_64)','(Macintosh; Intel Mac OS X 10_12_6)'
]chrome_version = 'Chrome/{}.0.{}.{}'.format(first_num, third_num, fourth_num)def get_ua():global os_type, chrome_versionreturn ' '.join(['Mozilla/5.0', random.choice(os_type), 'AppleWebKit/537.36','(KHTML, like Gecko)', chrome_version, 'Safari/537.36'])headers_index = {"Host": "www.mingxing.com",
}
headers_img = {"Referer": "http://www.mingxing.com/ziliao/index.html",
}root = 'starImages'
if not os.path.exists(root):os.makedirs(root)
s = requests.session()
for i in range(1, 194):url = r'http://www.mingxing.com/ziliao/index?&p={}'.format(i)headers_index['User-Agent'] = get_ua()response = s.get(url, headers=headers_index)html = etree.HTML(response.text)lis = html.xpath("//div[@class='page_starlist']//li")time.sleep(1)for li in lis:src = li.xpath(".//img/@src")[0]name = li.xpath(".//a/h3")[0].text.strip()print('Downloading {}'.format(name))headers_img['Referer'] = urlheaders_img['User-Agent'] = get_ua()img = s.get(src, headers=headers_img)folder = os.path.join(root, name)if not os.path.exists(folder):os.mkdir(folder)file = os.path.join(root, name, '{}.jpg'.format(name))with open(file, 'ab') as f:f.write(img.content)# time.sleep(0.2)img.close()

爬取结果，每个明星都建立一个文件夹，每个文件夹里有同名图片文件

2. 建立明星脸数据库

使用mtcnn与arcface对每一位明星的脸部进行提取并计算嵌入向量，通过分析向量之间的欧式距离即可判断相似度，主要代码如下，该函数将明星名字与人脸向量保存为本地文件facebank.pkl方便后续使用，具体涉及很多深度学习代码见原项目。

def get_facebank(path):names = []embeddings = []folders = os.listdir(path)with open("starlist.txt", 'w', encoding='utf-8') as f:for name in tqdm(folders):  # 迭代文件夹file = os.path.join(path, name, name + '.jpg')  # 获取图片文件名starimg = cv2.imdecode(np.fromfile(file, dtype=np.uint8), cv2.IMREAD_COLOR)if starimg is None: continue_, _, starface, embedding = face_model(starimg)  # 获取嵌入向量if len(embedding) != 1: continuef.write(name + '\n')names.append(name)embeddings.append(embedding[0])with open('facebank.pkl', 'wb') as f:pickle.dump((names, embeddings), f)  # 保存为本地文件方便使用return names, embeddings

3. 获取人脸对比数据库分析相似度

读取自己的照片，获取人脸与嵌入向量，与数据库的每一条向量进行比较，获取距离最近的向量即为最相似的人脸，获取两张图片的人脸进行对齐并展示。

def compare_embedding(embedding, facebank):"""单个人脸向量与数据库对比"""if len(facebank) == 0:return Noneembedding = np.array(embedding)facebank = np.array(facebank).squeeze(axis=1)diff = embedding - facebankdist = np.sum(np.square(diff), axis=1)min_idx = np.argmin(dist)  # 最小欧氏距离return min_idxdef find_similar_star(path, datapath='starImages'):myimg = cv2.imdecode(np.fromfile(path, dtype=np.uint8), cv2.IMREAD_COLOR)  # -1表示cv2.IMREAD_UNCHANGED_, _, myface, embedding = face_model(myimg)if len(myface) != 1:return None, None# 寻找最相似明星idx = compare_embedding(embedding[0], face_bank)starnname = names[idx]# 获取明星脸file = datapath + '\{}\{}.jpg'.format(starnname, starnname)starimg = cv2.imdecode(np.fromfile(file, dtype=np.uint8), cv2.IMREAD_COLOR)  # -1表示cv2.IMREAD_UNCHANGED_, _, starface, _ = face_model(starimg)# 人脸对比myface = myface[0]starface = starface[0]h, w = starface.shape[:2]myface = cv2.resize(myface, (w, h))# myface = cv2.GaussianBlur(myface, ksize=(15,15), sigmaX=0)result = np.hstack((myface, starface))return starnname, result

用我自己的图片测试结果如下

使用明星本人的图片测试结果如下

总结

可以看到arcface的人脸识别还是挺准的，相比那些小程序效果好的多而且最重要的是免费！！！

参考

mtcnn
insightface