Python轻松实现AI换脸

引言

在网上看到一篇基于python实现AI换脸的文章，觉得很好玩复现了一下。研究了几个小时后成功把迪丽热巴的脸换到关晓彤上，很好看；另外还把蔡徐坤的脸换到特朗普上，效果那是相当惊人(⊙ˍ⊙)

思路

思路比较简单，主要使用了旷视人工智能平台上的API实现。首先调用脸部识别API识别被换图和融合图，再调用脸部融合API换脸就完成了。整个过程中要用到request库调用API，json、simplejson和base64模块编码和解码，以及文件的读写操作
人脸识别API
脸部融合API

代码

import requests
import simplejson
import json
import base64


def detect_face(path, img):      # 脸部识别
    print("detecting %s..."%(img))
    url_detect = "https://api-cn.faceplusplus.com/facepp/v3/detect"    # 旷视Detect API
    data = {"api_key": "UmeTHKmayj5KFi6SCTt3-0rwc8XSx9k9",  
            "api_secret": "KlpfpUnIK_0FllNu3JbREJM3WKppz1OJ", "image_file":img , 
            "return_landmark": 2}      # 构造表单数据，用于post请求(api_key和api_secret要注册旷视人工智能平台后才有)
    files = {"image_file": open(path+img, "rb")}
    response = requests.post(url_detect, data=data, files=files)
    req_con = response.content.decode("utf-8")  # 将已编码的json字符解码为python对象
    req_dict = json.JSONDecoder().decode(req_con)
    this_json = simplejson.dumps(req_dict)  # 将python对象编码成json字符串
    this_json2 = simplejson.loads(this_json)  # 将已编码的json字符串解码为python对象
    faces = this_json2["faces"]
    # print(faces)
    rectangle = list0["face_rectangle"]
    return rectangle


def merge_face(path, img1, img2, img3, number):    # 人脸融合
    list0 = faces[0]
    df1 = detect_face(path, img1)
    df2 = detect_face(path, img2)
    rectangle1 = str(df1["top"]) + "," + str(df1["left"]) + "," + str(df1["width"]) + "," + str(df1["height"])
    rectangle2 = str(df2["top"]) + "," + str(df2["left"]) + "," + str(df2["width"]) + "," + str(df2["height"])
    url_merge = "https://api-cn.faceplusplus.com/imagepp/v1/mergeface" # 旷视Merge_face API
    f1 = open(path+img1, "rb")
    f1_64 = base64.b64encode(f1.read())
    f1.close()
    f2 = open(path+img2, "rb")
    f2_64 = base64.b64encode(f2.read())
    f2.close()
    data = {"api_key": "UmeTHKmayj5KFi6SCTt3-0rwc8XSx9k9", 
            "api_secret": "KlpfpUnIK_0FllNu3JbREJM3WKppz1OJ",
            "template_base64": f1_64, "template_rectangle": rectangle1,
            "merge_base64": f2_64, "merge_rectangle": rectangle2,
            "merge_rate": number}
    print("merging...")
    response = requests.post(url_merge, data=data)
    req_con = response.content.decode("utf-8")
    req_dict = json.JSONDecoder().decode(req_con)
    result = req_dict["result"]
    imgdata = base64.b64decode(result)
    file = open(path+img3, "wb")
    file.write(imgdata)
    file.close()


def main():
    path = input("输入图片路径：")
    image1 = input("输入被换图名称：")
    image2 = input("输入融合图名称：")
    image3 = input("输入换脸后的图片名称：")
    try:
        merge_face(path, image1, image2, image3, 100)  # 融合率为100
        print("converted successfully!")
    except:
        print("出现异常")
        import traceback
        traceback.print_exc()


main()

效果

经过测试发现：换脸效果和两张脸的朝向、脸型、大小、五官位置、是否闭眼有关（要得到不错的换脸图，条件还真有点苛刻），剔除了一些合成效果不好的图片后，保留了两组换脸效果还不错的图

还不错！
参考：
python实现AI换脸