tf23: “恶作剧” --人脸检测

MachineLP

发布于 2018-01-09 11:21:46

6690

发布于 2018-01-09 11:21:46

前面有一个帖《OpenCV检测场景内是否有移动物体》我用树莓派做了一个简单的Motion Detection，放在卫生间的，它会在我上大号时自动播放音乐。

我一个人租房，几个盆友周末时常会找我玩，他们觉得我做的Motion Detection很垃圾。于是我就想恶搞一下，用TensorFlow做一个“人脸识别”，在我上大号时播放音乐，如果是别人就播放《张震讲鬼故事》（@xingCI说放屁声更搞）。

我的任务的训练一个模型可以区分“我”和“其它人”的脸。注意，上面“人脸识别”我是加引号的，其实并不是真正的人脸识别，充其量就是个图像分类。如果你要使用真正的人脸识别，可以试试现成的库OpenFace+dlib《使用OpenFace进行人脸识别》。

有人已经把TensorFlow移植到了树莓派，项目地址tensorflow-on-raspberry-pi。

准备数据

本帖需要使用到两组数据：一组是包含我脸的图像，另一组包含其它人人脸的图像。

其它人人脸的收集

找一堆图片，只要不包含自己就行，然后使用OpenCV提取图像中的大脸。

提取图像中的人脸，我使用OpenCV，据说使用dlib效果更好。

other_peoples_faces.py：

import cv2
import os
import sys
 
IMAGE_DIR = '图片目录路径'
 
OUTPUT_DIR = './other_people'
if not os.path.exists(OUTPUT_DIR):
	os.makedirs(OUTPUT_DIR)
 
# http://blog.topspeedsnail.com/archives/10511
# wget https://raw.githubusercontent.com/opencv/opencv/master/data/haarcascades/haarcascade_frontalface_default.xml
face_haar = cv2.CascadeClassifier("haarcascade_frontalface_default.xml")
 
for (dirpath, dirnames, filenames) in os.walk(IMAGE_DIR):
	for filename in filenames:
		if filename.endswith('.jpg'):
			image_path = os.path.join(dirpath, filename)
			print('process: ', image_path)
			img = cv2.imread(image_path)
			
			gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
			faces = face_haar.detectMultiScale(gray_image, 1.3, 5)
			for face_x,face_y,face_w,face_h in faces:
				face = img[face_y:face_y+face_h, face_x:face_x+face_w]
 
				face = cv2.resize(face, (64, 64))
 
				cv2.imshow("img", face)
				cv2.imwrite(os.path.join(OUTPUT_DIR, filename), face)
 
			key = cv2.waitKey(30) & 0xff
			if key == 27:
				sys.exit(0)

4万多图片，我只提取了1万张脸，应该够使了。

图像大小 64×64

上面是OpenCV做的人脸检测，有了这个数据集又可以反过来训练TensorFlow版本的人脸检测。

斗大熊的脸

给自己拍照1万张，这是我一次拍照最多的一回。

import cv2
import os
import sys
 
OUTPUT_DIR = './my_faces'
if not os.path.exists(OUTPUT_DIR):
	os.makedirs(OUTPUT_DIR)
 
face_haar = cv2.CascadeClassifier("haarcascade_frontalface_default.xml")
 
cam = cv2.VideoCapture(0)
 
count = 0
while True:
	print(count)
	if count < 10000:
		_, img = cam.read()
 
		gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
		faces = face_haar.detectMultiScale(gray_image, 1.3, 5)
		for face_x,face_y,face_w,face_h in faces:
			face = img[face_y:face_y+face_h, face_x:face_x+face_w]
 
			face = cv2.resize(face, (64, 64))
			cv2.imshow('img', face)
			cv2.imwrite(os.path.join(OUTPUT_DIR, str(count)+'.jpg'), face)
			count += 1
		key = cv2.waitKey(30) & 0xff
		if key == 27:
			break
	else:
		break

在镜头前摇头晃脑、摆pose，戴眼镜、耳机，仰天45，写代码，呲牙咧嘴，玩手机。。。一定要多样化，直到拍1万张大脸。

训练模型

训练数据有了，下面开始训练。

import tensorflow as tf
import cv2
import numpy as np
import os
from sklearn.model_selection import train_test_split
import random
import sys
 
my_image_path = 'my_faces'
others_image_path = 'other_people'
 
image_data = []
label_data = []
 
def get_padding_size(image):
	h, w, _ = image.shape
	longest_edge = max(h, w)
	top, bottom, left, right = (0, 0, 0, 0)
	if h < longest_edge:
		dh = longest_edge - h
		top = dh // 2
		bottom = dh - top
	elif w < longest_edge:
		dw = longest_edge - w
		left = dw // 2
		right = dw - left
	else:
		pass
	return top, bottom, left, right
 
def read_data(img_path, image_h=64, image_w=64):
	for filename in os.listdir(img_path):
		if filename.endswith('.jpg'):
			filepath = os.path.join(img_path, filename)
			image = cv2.imread(filepath)
 
			top, bottom, left, right = get_padding_size(image)
			image_pad = cv2.copyMakeBorder(image, top , bottom, left, right, cv2.BORDER_CONSTANT, value=[0, 0, 0])
			image = cv2.resize(image_pad, (image_h, image_w))
 
			image_data.append(image)
			label_data.append(img_path)
 
read_data(others_image_path)
read_data(my_image_path)
 
image_data = np.array(image_data)
label_data = np.array([[0,1] if label == 'my_faces' else [1,0] for label in label_data])
 
train_x, test_x, train_y, test_y = train_test_split(image_data, label_data, test_size=0.05, random_state=random.randint(0, 100))
 
# image (height=64 width=64 channel=3)
train_x = train_x.reshape(train_x.shape[0], 64, 64, 3)
test_x = test_x.reshape(test_x.shape[0], 64, 64, 3)
 
# nomalize
train_x = train_x.astype('float32') / 255.0
test_x = test_x.astype('float32') / 255.0
 
print(len(train_x), len(train_y))
print(len(test_x), len(test_y))
 
#############################################################
batch_size = 128
num_batch = len(train_x) // batch_size
 
X = tf.placeholder(tf.float32, [None, 64, 64, 3])  # 图片大小64x64 channel=3
Y = tf.placeholder(tf.float32, [None, 2])
 
keep_prob_5 = tf.placeholder(tf.float32)
keep_prob_75 = tf.placeholder(tf.float32)
 
def panda_joke_cnn():
 
	W_c1 = tf.Variable(tf.random_normal([3, 3, 3, 32], stddev=0.01))
	b_c1 = tf.Variable(tf.random_normal([32]))
	conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(X, W_c1, strides=[1, 1, 1, 1], padding='SAME'), b_c1))
	conv1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
	conv1 = tf.nn.dropout(conv1, keep_prob_5)
 
	W_c2 = tf.Variable(tf.random_normal([3, 3, 32, 64], stddev=0.01))
	b_c2 = tf.Variable(tf.random_normal([64]))
	conv2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv1, W_c2, strides=[1, 1, 1, 1], padding='SAME'), b_c2))
	conv2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
	conv2 = tf.nn.dropout(conv2, keep_prob_5)
 
	W_c3 = tf.Variable(tf.random_normal([3, 3, 64, 64], stddev=0.01))
	b_c3 = tf.Variable(tf.random_normal([64]))
	conv3 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(conv2, W_c3, strides=[1, 1, 1, 1], padding='SAME'), b_c3))
	conv3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
	conv3 = tf.nn.dropout(conv3, keep_prob_5)
 
	# Fully connected layer
	W_d = tf.Variable(tf.random_normal([8*16*32, 512], stddev=0.01))
	b_d = tf.Variable(tf.random_normal([512]))
	dense = tf.reshape(conv3, [-1, W_d.get_shape().as_list()[0]])
	dense = tf.nn.relu(tf.add(tf.matmul(dense, W_d), b_d))
	dense = tf.nn.dropout(dense, keep_prob_75)
 
	W_out = tf.Variable(tf.random_normal([512, 2], stddev=0.01))
	b_out = tf.Variable(tf.random_normal([2]))
	out = tf.add(tf.matmul(dense, W_out), b_out)
	return out
 
def train_cnn():
	output = panda_joke_cnn()
 
	loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(output, Y))
	optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
 
	accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(output, 1), tf.argmax(Y, 1)), tf.float32))
 
	tf.summary.scalar("loss", loss)
	tf.summary.scalar("accuracy", accuracy)
	merged_summary_op = tf.summary.merge_all()
 
	saver = tf.train.Saver()
	with tf.Session() as sess:
		sess.run(tf.global_variables_initializer())
 
		summary_writer = tf.summary.FileWriter('./log', graph=tf.get_default_graph())
 
		for e in range(50):
			for i in range(num_batch):
				batch_x = train_x[i*batch_size : (i+1)*batch_size]
				batch_y = train_y[i*batch_size : (i+1)*batch_size]
				_, loss_, summary = sess.run([optimizer, loss, merged_summary_op], feed_dict={X: batch_x, Y: batch_y, keep_prob_5:0.5, keep_prob_75: 0.75})
 
				summary_writer.add_summary(summary, e*num_batch+i)
				print(e*num_batch+i, loss_)
 
				if (e*num_batch+i) % 100 == 0:
					acc = accuracy.eval({X: test_x, Y: test_y, keep_prob_5:1.0, keep_prob_75: 1.0})
					print(e*num_batch+i, acc)
					# save model
					if acc > 0.98:
						saver.save(sess, "i_am_a_joke.model", global_step=e*num_batch+i)
						sys.exit(0)
 
train_cnn()

准确率曲线：

下面要做的就是在树莓派上使用模型，代码示例：

output = panda_joke_cnn()
predict = tf.argmax(output, 1)
 
saver = tf.train.Saver()
sess = tf.Session()
saver.restore(sess, tf.train.latest_checkpoint('.'))
 
def is_my_face(image):
	res = sess.run(predict, feed_dict={X: [image/255.0], keep_prob_5:1.0, keep_prob_75: 1.0})
	if res[0] == 1:
		return True
	else:
		return False
 
face_haar = cv2.CascadeClassifier("haarcascade_frontalface_default.xml")
cam = cv2.VideoCapture(0)
 
while True:
	_, img = cam.read()
	gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
	faces = face_haar.detectMultiScale(gray_image, 1.3, 5)
	for face_x,face_y,face_w,face_h in faces:
		face = img[face_y:face_y+face_h, face_x:face_x+face_w]
 
		face = cv2.resize(face, (64, 64))
 
		print(is_my_face(face))
 
		cv2.imshow('img', face)
		key = cv2.waitKey(30) & 0xff
		if key == 27:
			sys.exit(0)
 
sess.close()

总结：占用内存100多M，准确率还凑合，先用着。

Share the post "TensorFlow练习23: “恶作剧”"

本文参与腾讯云自媒体分享计划，分享自作者个人站点/博客。

原始发表：2017年01月02日，如有侵权请联系 cloudcommunity@tencent.com 删除

机器学习