基于Python与TensorFlow的人脸识别验证实现

核心代码结构解析

# -*- coding: utf-8 -*-
# __author__="maple"
"""
       ┏┓   ┏┓
      ┏┛┻━━━┛┻┓
      ┃   ☃   ┃
      ┃ ┳┛ ┗┳ ┃
      ┃   ┻   ┃
      ┗━┓   ┏━┛
        ┃   ┗━━━┓
        ┃ 神兽保佑  ┣┓
        ┃　永无BUG！  ┏┛
        ┗┓┓┏━┳┓┏┛
         ┃┫┫ ┃┫┫
         ┗┻┛ ┗┻┛
"""
import base64
import cv2
import time
from io import BytesIO
from tensorflow import keras
from PIL import Image
from pymongo import MongoClient
import tensorflow as tf
import face_recognition
import numpy as np
#mongodb连接
conn = MongoClient('mongodb://root:123@localhost:27017/')
db = conn.myface #连接mydb数据库，没有则自动创建
user_face = db.user_face #使用test_set集合，没有则自动创建
face_images = db.face_images


lables = []
datas = []
INPUT_NODE = 128
LATER1_NODE = 200
OUTPUT_NODE = 0
TRAIN_DATA_SIZE = 0
TEST_DATA_SIZE = 0


def generateds():
  get_out_put_node()
  train_x, train_y, test_x, test_y = np.array(datas),np.array(lables),np.array(datas),np.array(lables)
  return train_x, train_y, test_x, test_y

def get_out_put_node():
  for item in face_images.find():
    lables.append(item['user_id'])
    datas.append(item['face_encoding'])
  OUTPUT_NODE = len(set(lables))
  TRAIN_DATA_SIZE = len(lables)
  TEST_DATA_SIZE = len(lables)
  return OUTPUT_NODE, TRAIN_DATA_SIZE, TEST_DATA_SIZE

# 验证脸部信息
def predict_image(image):
  model = tf.keras.models.load_model('face_model.h5',compile=False)
  face_encode = face_recognition.face_encodings(image)
  result = []
  for j in range(len(face_encode)):
    predictions1 = model.predict(np.array(face_encode[j]).reshape(1, 128))
    print(predictions1)
    if np.max(predictions1[0]) > 0.90:
      print(np.argmax(predictions1[0]).dtype)
      pred_user = user_face.find_one({'id': int(np.argmax(predictions1[0]))})
      print('第%d张脸是%s' % (j+1, pred_user['user_name']))
      result.append(pred_user['user_name'])
  return result

# 保存脸部信息
def save_face(pic_path,uid):
  image = face_recognition.load_image_file(pic_path)
  face_encode = face_recognition.face_encodings(image)
  print(face_encode[0].shape)
  if(len(face_encode) == 1):
    face_image = {
      'user_id': uid,
      'face_encoding':face_encode[0].tolist()
    }
    face_images.insert_one(face_image)

# 训练脸部信息
def train_face():
  train_x, train_y, test_x, test_y = generateds()
  dataset = tf.data.Dataset.from_tensor_slices((train_x, train_y))
  dataset = dataset.batch(32)
  dataset = dataset.repeat()
  OUTPUT_NODE, TRAIN_DATA_SIZE, TEST_DATA_SIZE = get_out_put_node()
  model = keras.Sequential([
    keras.layers.Dense(128, activation=tf.nn.relu),
    keras.layers.Dense(128, activation=tf.nn.relu),
    keras.layers.Dense(OUTPUT_NODE, activation=tf.nn.softmax)
  ])

  model.compile(optimizer=tf.compat.v1.train.AdamOptimizer(),
        loss='sparse_categorical_crossentropy',
        metrics=['accuracy'])
  steps_per_epoch = 30
  if steps_per_epoch > len(train_x):
    steps_per_epoch = len(train_x)
  model.fit(dataset, epochs=10, steps_per_epoch=steps_per_epoch)

  model.save('face_model.h5')



def register_face(user):
  if user_face.find({"user_name": user}).count() > 0:
    print("用户已存在")
    return
  video_capture=cv2.VideoCapture(0)
  # 在MongoDB中使用sort()方法对数据进行排序，sort()方法可以通过参数指定排序的字段，并使用 1 和 -1 来指定排序的方式，其中 1 为升序，-1为降序。
  finds = user_face.find().sort([("id", -1)]).limit(1)
  uid = 0
  if finds.count() > 0:
    uid = finds[0]['id'] + 1
  print(uid)
  user_info = {
    'id': uid,
    'user_name': user,
    'create_time': time.time(),
    'update_time': time.time()
  }
  user_face.insert_one(user_info)

  while 1:
    # 获取一帧视频
    ret, frame = video_capture.read()
    # 窗口显示
    cv2.imshow('Video',frame)
    # 调整角度后连续拍5张图片
    if cv2.waitKey(1) & 0xFF == ord('q'):
      for i in range(1,6):
        cv2.imwrite('Myface{}.jpg'.format(i), frame)
        with open('Myface{}.jpg'.format(i),"rb")as f:
          img=f.read()
          img_data = BytesIO(img)
          im = Image.open(img_data)
          im = im.convert('RGB')
          imgArray = np.array(im)
          faces = face_recognition.face_locations(imgArray)
          save_face('Myface{}.jpg'.format(i),uid)
      break

  train_face()
  video_capture.release()
  cv2.destroyAllWindows()


def rec_face():
  video_capture = cv2.VideoCapture(0)
  while 1:
    # 获取一帧视频
    ret, frame = video_capture.read()
    # 窗口显示
    cv2.imshow('Video',frame)
    # 验证人脸的5照片
    if cv2.waitKey(1) & 0xFF == ord('q'):
      for i in range(1,6):
        cv2.imwrite('recface{}.jpg'.format(i), frame)
      break

  res = []
  for i in range(1, 6):
    with open('recface{}.jpg'.format(i),"rb")as f:
      img=f.read()
      img_data = BytesIO(img)
      im = Image.open(img_data)
      im = im.convert('RGB')
      imgArray = np.array(im)
      predict = predict_image(imgArray)
      if predict:
        res.extend(predict)

  b = set(res) # {2, 3}
  if len(b) == 1 and len(res) >= 3:
    print(" 验证成功")
  else:
    print(" 验证失败")

if __name__ == '__main__':
  register_face("maple")
  rec_face()