基于CNN卷积神经网络的TensorFlow+Keras深度学习的人脸识别

本站内容均来自兴趣收集,如不慎侵害的您的相关权益,请留言告知,我们将尽快删除.谢谢.

基于CNN卷积神经网络的TensorFlow+Keras深度学习的人脸识别

在上一篇博客中，利用CNN卷积神经网络的TensorFlow+Keras深度学习搭建了人脸模型：

基于CNN卷积神经网络的TensorFlow+Keras深度学习搭建人脸模型

本篇博客将继续利用CNN卷积神经网络的TensorFlow+Keras深度学习
实现人脸识别

项目实现效果

PS：项目地址在最后会开源

本项目使用TensorFlow-GPU
进行训练：需要提前搭建好CUDA环境
具体可以参考本文：TensorFlow-GPU-2.4.1与CUDA安装教程

模型数据

嵌入模型

Model: "embedding"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_image (InputLayer)     [(None, 100, 100, 3)]     0         
_________________________________________________________________
conv2d (Conv2D)              (None, 91, 91, 64)        19264     
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 46, 46, 64)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 40, 40, 128)       401536    
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 20, 20, 128)       0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 17, 17, 128)       262272    
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 9, 9, 128)         0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 6, 6, 256)         524544    
_________________________________________________________________
flatten (Flatten)            (None, 9216)              0         
_________________________________________________________________
dense (Dense)                (None, 4096)              37752832  
=================================================================
Total params: 38,960,448
Trainable params: 38,960,448
Non-trainable params: 0
_________________________________________________________________

CNN神经网络模型

Model: "SiameseNetWork"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_image (InputLayer)        [(None, 100, 100, 3) 0                                            
__________________________________________________________________________________________________
validation_img (InputLayer)     [(None, 100, 100, 3) 0                                            
__________________________________________________________________________________________________
embedding (Functional)          (None, 4096)         38960448    input_image[0][0]                
                                                                 validation_img[0][0]             
__________________________________________________________________________________________________
distance (L1Dist)               (None, 4096)         0           embedding[4][0]                  
                                                                 embedding[5][0]                  
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 1)            4097        distance[0][0]                   
==================================================================================================
Total params: 38,964,545
Trainable params: 38,964,545
Non-trainable params: 0
__________________________________________________________________________________________________

项目概述

项目运行流程

1. 收集人脸数据—设置数据的路径并对数据集预处理

2. 构建训练模型——搭建深度神经网络

3. 深度训练人脸数据——CNN卷积神经网络+TensorFlow+Keras

4. 搭建人脸识别APP——OpenCV+Kivy.APP

核心环境配置

Python == 3.9.0
labelme == 5.0.1
tensorflow -gpu == 2.7.0 （CUDA11.2）
opencv-python == 4.0.1
Kivy == 2.1.0
albumentations == 0.7.12

项目核心代码详解

本项目用的到野生人脸数据集下载地址：深度学习人脸训练数据集

本项目基于《Siamese Neural Networks for One-shot Image Recognition》这篇论文为理论基础：Siamese Neural Networks for One-shot Image Recognition

核心代码

引入的核心库文件：

import cv2
import numpy as np
from matplotlib import pyplot as plt
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Layer, Conv2D, Dense, MaxPooling2D, Input, Flatten
import tensorflow as tf

加入GPU内存增长限制—防止爆显存

gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus: 
    tf.config.experimental.set_memory_growth(gpu, True)

设置数据集目录

POS_PATH = os.path.join('data', 'positive')
NEG_PATH = os.path.join('data', 'negative')
ANC_PATH = os.path.join('data', 'anchor')
os.makedirs(POS_PATH)
os.makedirs(NEG_PATH)
os.makedirs(ANC_PATH)
# 导入野生数据集
for directory in os.listdir('666'):
    for file in os.listdir(os.path.join('666', directory)):
        EX_PATH = os.path.join('666', directory, file)
        NEW_PATH = os.path.join(NEG_PATH, file)
        os.replace(EX_PATH, NEW_PATH)

收集人脸识别数据——UUID格式命名

cap = cv2.VideoCapture(0)
while cap.isOpened(): 
    ret, frame = cap.read()
   
    # 裁剪图像大小250x250px
    frame = frame[120:120+250,200:200+250, :]
    
    # 收集人脸数据——正面清晰的数据集
    if cv2.waitKey(1) & 0XFF == ord('a'):
        # 对数据进行UUID命名
        imgname = os.path.join(ANC_PATH, '{}.jpg'.format(uuid.uuid1()))
        # 写入并保存数据
        cv2.imwrite(imgname, frame)
    
    # 收集数据集——侧脸斜脸的数据集（可以较模糊）
    if cv2.waitKey(1) & 0XFF == ord('p'):
        # 对数据进行UUID命名
        imgname = os.path.join(POS_PATH, '{}.jpg'.format(uuid.uuid1()))
        # 写入并保存数据
        cv2.imwrite(imgname, frame)
    
    cv2.imshow('Image Collection', frame)
    
    if cv2.waitKey(1) & 0XFF == ord('q'):
        break
        
# 释放摄像头资源
cap.release()
cv2.destroyAllWindows()

创建标签化数据集

positives = tf.data.Dataset.zip((anchor, positive, tf.data.Dataset.from_tensor_slices(tf.ones(len(anchor)))))
negatives = tf.data.Dataset.zip((anchor, negative, tf.data.Dataset.from_tensor_slices(tf.zeros(len(anchor)))))
data = positives.concatenate(negatives)
samples = data.as_numpy_iterator()
exampple = samples.next()

构建训练和测试数据的分区

def preprocess_twin(input_img, validation_img, label):
    return(preprocess(input_img), preprocess(validation_img), label)
res = preprocess_twin(*exampple)
data = data.map(preprocess_twin)
data = data.cache()
data = data.shuffle(buffer_size=1024)
train_data = data.take(round(len(data)*.7))
train_data = train_data.batch(16)
train_data = train_data.prefetch(8)
test_data = data.skip(round(len(data)*.7))
test_data = test_data.take(round(len(data)*.3))
test_data = test_data.batch(16)
test_data = test_data.prefetch(8)``

创建模型

inp = Input(shape=(100,100,3), name='input_image')
c1 = Conv2D(64, (10,10), activation='relu')(inp)
m1 = MaxPooling2D(64, (2,2), padding='same')(c1)
c2 = Conv2D(128, (7,7), activation='relu')(m1)
m2 = MaxPooling2D(64, (2,2), padding='same')(c2)
c3 = Conv2D(128, (4,4), activation='relu')(m2)
m3 = MaxPooling2D(64, (2,2), padding='same')(c3)
c4 = Conv2D(256, (4,4), activation='relu')(m3)
f1 = Flatten()(c4)
d1 = Dense(4096, activation='sigmoid')(f1)
mod = Model(inputs=[inp], outputs=[d1], name='embedding')
mod.summary()

def make_embedding(): 
    inp = Input(shape=(100,100,3), name='input_image')
    # 第一层卷积
    c1 = Conv2D(64, (10,10), activation='relu')(inp)
    m1 = MaxPooling2D(64, (2,2), padding='same')(c1)
    
    # 第二层卷积
    c2 = Conv2D(128, (7,7), activation='relu')(m1)
    m2 = MaxPooling2D(64, (2,2), padding='same')(c2)
    
    # 第三层卷积 
    c3 = Conv2D(128, (4,4), activation='relu')(m2)
    m3 = MaxPooling2D(64, (2,2), padding='same')(c3)
    
    # 最终卷积
    c4 = Conv2D(256, (4,4), activation='relu')(m3)
    f1 = Flatten()(c4)
    d1 = Dense(4096, activation='sigmoid')(f1)
    
    
    return Model(inputs=[inp], outputs=[d1], name='embedding')
    embedding = make_embedding()

构建距离层

# L1距离层
class L1Dist(Layer):
    
    # 初始化方法
    def __init__(self, **kwargs):
        super().__init__()
       
    # 数据相似度计算
    def call(self, input_embedding, validation_embedding):
        return tf.math.abs(input_embedding - validation_embedding)
        l1 = L1Dist()
        l1(anchor_embedding, validation_embedding)

构建神经网络模型

input_image = Input(name='input_img', shape=(100,100,3))
validation_image = Input(name='validation_img', shape=(100,100,3))
inp_embedding = embedding(input_image)
val_embedding = embedding(validation_image)
siamese_layer = L1Dist()
distances = siamese_layer(inp_embedding, val_embedding)
classifier = Dense(1, activation='sigmoid')(distances)
siamese_network = Model(inputs=[input_image, validation_image], outputs=classifier, name='SiameseNetwork')

深度训练模型

搭建损失值和优化器

binary_cross_loss = tf.losses.BinaryCrossentropy()
opt = tf.keras.optimizers.Adam(1e-4) # 0.0001

设置训练检查点

checkpoint_dir = './training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, 'ckpt')
checkpoint = tf.train.Checkpoint(opt=opt, siamese_model=siamese_model)

设置训练batch

test_batch = train_data.as_numpy_iterator()
batch_1 = test_batch.next()
X = batch_1[:2]
y = batch_1[2]
@tf.function
def train_step(batch):
    
    # 日志记录
    with tf.GradientTape() as tape:     
        # 获得人脸数据
        X = batch[:2]
        # 获得标签
        y = batch[2]
        
        # yhat的值向上传递
        yhat = siamese_model(X, training=True)
        # 计算损失值
        loss = binary_cross_loss(y, yhat)
    print(loss)
        
    # 计算渐变值
    grad = tape.gradient(loss, siamese_model.trainable_variables)
    
    # 计算更新的权重传递给模型
    opt.apply_gradients(zip(grad, siamese_model.trainable_variables))
    
    # 返回损失值
    return loss

设置训练循环

def train(data, EPOCHS):
    # Loop through epochs
    for epoch in range(1, EPOCHS+1):
        print('
 Epoch {}/{}'.format(epoch, EPOCHS))
        progbar = tf.keras.utils.Progbar(len(data))
        
        # Loop through each batch
        for idx, batch in enumerate(data):
            # Run train step here
            train_step(batch)
            progbar.update(idx+1)
        
        # Save checkpoints
        if epoch % 10 == 0: 
            checkpoint.save(file_prefix=checkpoint_prefix)

开始训练

EPOCHS = 50000
train(train_data, EPOCHS)

保存模型

siamese_model.save('siamesemodel.h5')

加载模型

model = tf.keras.models.load_model('siamesemodel.h5', 
                                   custom_objects={
 'L1Dist':L1Dist, 'BinaryCrossentropy':tf.losses.BinaryCrossentropy})

测试模型识别效果

cap = cv2.VideoCapture(0)
while cap.isOpened():
    ret, frame = cap.read()
    frame = frame[120:120+250,200:200+250, :]
    
    cv2.imshow('Verification', frame)
    
   
    if cv2.waitKey(10) & 0xFF == ord('v'):  
        cv2.imwrite(os.path.join('application_data', 'input_image', 'input_image.jpg'), frame)
        results, verified = verify(model, 0.9, 0.7)
        print(verified)
    
    if cv2.waitKey(10) & 0xFF == ord('q'):
        break
cap.release()
cv2.destroyAllWindows()

人脸识别APP—窗口UI

# Coding BIGBOSSyifi
# Datatime:2022/4/27 22:07
# Filename:FaceAPP.py
# Toolby: PyCharm
# 本篇代码实现功能：加载模型通过摄像头进行验证 代码51可修改模型路径
from kivy.app import App
from kivy.uix.boxlayout import BoxLayout
from kivy.uix.image import Image
from kivy.uix.button import Button
from kivy.uix.label import Label
from kivy.clock import Clock 
from kivy.graphics.texture import Texture
from kivy.logger import Logger
import cv2
import tensorflow as tf
from tensorflow.keras.layers import Layer
import os
import numpy as np
# 向命运妥协法(CPU)：
#os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
#os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
class L1Dist(Layer):
    def __init__(self, **kwargs):
        super().__init__()
    # 相似性计算：
    def call(self, input_embedding, validation_embedding):
        return tf.math.abs(input_embedding - validation_embedding)
# 构建APP布局：
class CamApp(App):
    def build(self):
        # 主界面布局：
        self.web_cam = Image(size_hint=(1, .8))
        self.button = Button(text="Start Verify", on_press=self.verify, size_hint=(1, .1))
        self.verification_label = Label(text="Verification Uninitiated...", size_hint=(1, .1))
        # 添加按键功能
        layout = BoxLayout(orientation='vertical')
        layout.add_widget(self.web_cam)
        layout.add_widget(self.button)
        layout.add_widget(self.verification_label)
        # 加载tensorflow/keras模型
        self.model = tf.keras.models.load_model('siamesemodelPRO.h5', custom_objects={
 'L1Dist': L1Dist})
        # 设置cv2摄像捕捉
        self.capture = cv2.VideoCapture(0)
        Clock.schedule_interval(self.update, 1.0 / 33.0)
        return layout
    # 连续获取摄像头图像
    def update(self, *args):
        # 读取cv2的框架：
        ret, frame = self.capture.read()
        frame = frame[120:120 + 250, 200:200 + 250, :]      # 对摄像捕捉图像裁剪
        # 翻转水平并将图像转换为纹理图像
        buf = cv2.flip(frame, 0).tostring()
        img_texture = Texture.create(size=(frame.shape[1], frame.shape[0]), colorfmt='bgr')
        img_texture.blit_buffer(buf, colorfmt='bgr', bufferfmt='ubyte')
        self.web_cam.texture = img_texture
    # 将图像从文件和转换器转换为100x100px
    def preprocess(self, file_path):
        # 读取路径图片
        byte_img = tf.io.read_file(file_path)
        # 加载路径图片
        img = tf.io.decode_jpeg(byte_img)
        # 预处理步骤-将图像大小调整为100x100x3 (3通道)
        img = tf.image.resize(img, (100, 100))
        # 将图像缩放到0到1之间
        img = img / 255.0
        # Return image
        return img
    # 验证人脸图像
    def verify(self, *args):
        # 指定阈值
        detection_threshold = 0.99
        verification_threshold = 0.8         # 近似值设置
        SAVE_PATH = os.path.join('application_data', 'input_image', 'input_image.jpg')
        ret, frame = self.capture.read()
        frame = frame[120:120 + 250, 200:200 + 250, :]
        cv2.imwrite(SAVE_PATH, frame)
        # 生成结果数组
        results = []
        for image in os.listdir(os.path.join('application_data', 'verification_images')):
            input_img = self.preprocess(os.path.join('application_data', 'input_image', 'input_image.jpg'))
            validation_img = self.preprocess(os.path.join('application_data', 'verification_images', image))
            # 对模型进行预测(验证)
            result = self.model.predict(list(np.expand_dims([input_img, validation_img], axis=1)))
            results.append(result)
        # 检测阈值：高于该阈值的预测被认为是正的指标
        detection = np.sum(np.array(results) > detection_threshold)
        # 验证阈值：阳性预测/总阳性样本的比例
        verification = detection / len(os.listdir(os.path.join('application_data', 'verification_images')))
        verified = verification > verification_threshold
        # 设置APP文本
        self.verification_label.text = 'Verified' if verified == True else 'Unverified'
        # 输出验证结果
        Logger.info(results)
        Logger.info(detection)
        Logger.info(verification)
        Logger.info(verified)
        return results, verified
if __name__ == '__main__':
    CamApp().run()