AD
新智元原创
来源：Reddit、GitHub
编辑: 金磊
【新智元导读】人脸特征点检测是人脸检测过程中的一个重要环节。以往我们采用的方法是OpenCV或者Dlib，虽然Dlib优于OpenCV，但是检测出的68个点并没有覆盖额头区域。Reddit一位网友便在此基础上做了进一步研究，能够检测出81个面部特征点，使得准确度有所提高。
或许，这就是你需要的人脸特征点检测方法。
人脸特征点检测(Facial landmark detection)是人脸检测过程中的一个重要环节。是在人脸检测的基础上进行的，对人脸上的特征点例如嘴角、眼角等进行定位。
近日，Reddit一位网友po出一个帖子，表示想与社区同胞们分享自己的一点研究成果：
其主要的工作就是在人脸检测Dlib库68个特征点的基础上，增加了13个特征点(共81个) ，使得头部检测和图像操作更加精确。
现在来看一下demo：
demo视频链接：
https://www.youtube.com/watch?v=mDJrASIB1T0
81个特征点，人脸特征点检测更加精准
以往我们在做人脸特征点检测的时候，通常会用OpenCV来进行操作。
但自从人脸检测Dlib库问世，网友们纷纷表示：好用！Dlib≥OpenCV！Dlib具有更多的人脸识别模型，可以检测脸部68甚至更多的特征点。
我们来看一下Dlib的效果：
Dlib人脸特征点检测效果图
那么这68个特征点又是如何分布的呢？请看下面这张“面相图”：
人脸68个特征点分布
但无论是效果图和“面相图”，我们都可以发现在额头区域是没有分布特征点的。
于是，网友便提出了一个特征点能够覆盖额头区域的模型。
该模型是一个自定义形状预测模型，在经过训练后，可以找到任何给定图像中的81个面部特征点。
它的训练方法类似于Dlib的68个面部特征点形状预测器。只是在原有的68个特征点的基础上，在额头区域增加了13个点。这就使得头部的检测，以及用于需要沿着头部顶部的点的图像操作更加精准。
81个特征点效果图
这13个额外的特征点提取的方法，是根据该博主之前的工作完成的。
GitHub地址：
https://github.com/codeniko/eos
该博主继续使用Surrey Face Model，并记下了他认为适合他工作的13个点，并做了一些细节的修改。
当然，博主还慷慨的分享了训练的代码：
1#!/usr/bin/python
2# The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
3#
4# This example program shows how to use dlib's implementation of the paper:
5# One Millisecond Face Alignment with an Ensemble of Regression Trees by
6# Vahid Kazemi and Josephine Sullivan, CVPR 2014
7#
8# In particular, we will train a face landmarking model based on a small
9# dataset and then evaluate it. If you want to visualize the output of the
10# trained model on some images then you can run the
11# face_landmark_detection.py example program with predictor.dat as the input
12# model.
13#
14# It should also be noted that this kind of model, while often used for face
15# landmarking, is quite general and can be used for a variety of shape
16# prediction tasks. But here we demonstrate it only on a simple face
17# landmarking task.
18#
19# COMPILING/INSTALLING THE DLIB PYTHON INTERFACE
20# You can install dlib using the command:
21# pip install dlib
22#
23# Alternatively, if you want to compile dlib yourself then go into the dlib
24# root folder and run:
25# python setup.py install
26#
27# Compiling dlib should work on any operating system so long as you have
28# CMake installed. On Ubuntu, this can be done easily by running the
29# command:
30# sudo apt-get install cmake
31#
32# Also note that this example requires Numpy which can be installed
33# via the command:
34# pip install numpy
35
36import os
37import sys
38import glob
39
40import dlib
41
42# In this example we are going to train a face detector based on the small
43# faces dataset in the examples/faces directory. This means you need to supply
44# the path to this faces folder as a command line argument so we will know
45# where it is.
46if len(sys.argv) != 2:
47 print(
48 "Give the path to the examples/faces directory as the argument to this "
49 "program. For example, if you are in the python_examples folder then "
50 "execute this program by running:\n"
51 " ./train_shape_predictor.py ../examples/faces")
52 exit()
53faces_folder = sys.argv[1]
54
55options = dlib.shape_predictor_training_options()
56# Now make the object responsible for training the model.
57# This algorithm has a bunch of parameters you can mess with. The
58# documentation for the shape_predictor_trainer explains all of them.
59# You should also read Kazemi's paper which explains all the parameters
60# in great detail. However, here I'm just setting three of them
61# differently than their default values. I'm doing this because we
62# have a very small dataset. In particular, setting the oversampling
63# to a high amount (300) effectively boosts the training set size, so
64# that helps this example.
65options.oversampling_amount = 300
66# I'm also reducing the capacity of the model by explicitly increasing
67# the regularization (making nu smaller) and by using trees with
68# smaller depths.
69options.nu = 0.05
70options.tree_depth = 2
71options.be_verbose = True
72
73# dlib.train_shape_predictor() does the actual training. It will save the
74# final predictor to predictor.dat. The input is an XML file that lists the
75# images in the training dataset and also contains the positions of the face
76# parts.
77training_xml_path = os.path.join(faces_folder, "training_with_face_landmarks.xml")
78dlib.train_shape_predictor(training_xml_path, "predictor.dat", options)
79
80# Now that we have a model we can test it. dlib.test_shape_predictor()
81# measures the average distance between a face landmark output by the
82# shape_predictor and where it should be according to the truth data.
83print("\nTraining accuracy: {}".format(
84 dlib.test_shape_predictor(training_xml_path, "predictor.dat")))
85# The real test is to see how well it does on data it wasn't trained on. We
86# trained it on a very small dataset so the accuracy is not extremely high, but
87# it's still doing quite good. Moreover, if you train it on one of the large
88# face landmarking datasets you will obtain state-of-the-art results, as shown
89# in the Kazemi paper.
90testing_xml_path = os.path.join(faces_folder, "testing_with_face_landmarks.xml")
91print("Testing accuracy: {}".format(
92 dlib.test_shape_predictor(testing_xml_path, "predictor.dat")))
93
94# Now let's use it as you would in a normal application. First we will load it
95# from disk. We also need to load a face detector to provide the initial
96# estimate of the facial location.
97predictor = dlib.shape_predictor("predictor.dat")
98detector = dlib.get_frontal_face_detector()
99
100# Now let's run the detector and shape_predictor over the images in the faces
101# folder and display the results.
102print("Showing detections and predictions on the images in the faces folder...")
103win = dlib.image_window()
104for f in glob.glob(os.path.join(faces_folder, "*.jpg")):
105 print("Processing file: {}".format(f))
106 img = dlib.load_rgb_image(f)
107
108 win.clear_overlay()
109 win.set_image(img)
110
111 # Ask the detector to find the bounding boxes of each face. The 1 in the
112 # second argument indicates that we should upsample the image 1 time. This
113 # will make everything bigger and allow us to detect more faces.
114 dets = detector(img, 1)
115 print("Number of faces detected: {}".format(len(dets)))
116 for k, d in enumerate(dets):
117 print("Detection {}: Left: {} Top: {} Right: {} Bottom: {}".format(
118 k, d.left(), d.top(), d.right(), d.bottom()))
119 # Get the landmarks/parts for the face in box d.
120 shape = predictor(img, d)
121 print("Part 0: {}, Part 1: {} ...".format(shape.part(0),
122 shape.part(1)))
123 # Draw the face landmarks on the screen.
124 win.add_overlay(shape)
125
126 win.add_overlay(dets)
127 dlib.hit_enter_to_continue()
有需要的小伙伴们，快来试试这个模型吧！
参考链接：
GitHub：
https://github.com/codeniko/shape_predictor_81_face_landmarks
Reddit：
https://www.reddit.com/r/MachineLearning/comments/b20b9i/p_i_trained_a_face_predictor_that_detects_fulls/
Youtube：
https://www.youtube.com/watch?v=mDJrASIB1T0
新智元春季招聘开启，一起弄潮AI之巅！
岗位详情请戳：
【春招英雄贴】新智元呼召智士主笔，2019勇闯AI之巅！
【2019新智元 AI 技术峰会倒计时9天】
2019年的3月27日，新智元再汇AI之力，在北京泰富酒店举办AI开年盛典——2019新智元AI技术峰会。峰会以“智能云?芯世界“为主题，聚焦智能云和AI芯片的发展，重塑未来AI世界格局。
同时，新智元将在峰会现场权威发布若干AI白皮书，聚焦产业链的创新活跃，评述华人AI学者的影响力，助力中国在世界级的AI竞争中实现超越。
购票二维码
活动行购票链接：http://hdxu.cn/9Lb5U
点击文末“阅读原文”，马上参会！
本文首发于微信公众号：新智元。文章内容属作者个人观点，不代表和讯网立场。投资者据此操作，风险请自担。
