经典文章（Deep Convolutional Network Cascade for Facial Point Detection）一文的实现

    总之，经过了长时间的前期准备，我开辟了自己的新的研究方向——人脸检测和人脸配准。光说不练假把式，读了那么多文章，我能够详细了解人脸配准的整体思想，但是没有动手实践是没有任何用处的，于是我在自己的笔记本上仿真一篇文章，Deep Convolutional Network Cascade for Facial Point Detection。

这篇文章主要思想框架如下：

主要仿真的是Deep CNN F1网络，整个框架太复杂，于是我只是实现基本的第一层。

1）首先下载本文文章的数据http://mmlab.ie.cuhk.edu.hk/archive/CNN_FacePoint.htm，这个项目主页中有两个可执行文件，不过没啥用，运行不起来。

2）第二，虽然有了数据库但是不一定能使用，需要转化为HDF5文件才能使用。

import sys 
sys.path.append(r'C:\Users\改路径\Desktop\CNN-Face\dataset')#这两行是为了加载utils模块
import utils
import os
from os.path import join, exists
import cv2  #这个需要配置opencv,然后调用
import numpy as np
import h5py
from utils import flip, getDataFromTxt, rotate, logger, shuffle_in_unison_scary, processImage, createDir


TRAIN = 'C:/Users/改/Desktop/train'
OUTPUT = 'C:/Users/wenjie/Desktop/train'
if not exists(OUTPUT): os.mkdir(OUTPUT)
assert(exists(TRAIN) and exists(OUTPUT))


def generate_hdf5_data(filelist, output, fname, argument=False):
    data = getDataFromTxt(filelist)
    F_imgs = []
    F_landmarks = []
    EN_imgs = []
    EN_landmarks = []
    NM_imgs = []
    NM_landmarks = []
    for (imgPath, bbox, landmarkGt) in data:    
        img = cv2.imread(imgPath, cv2.CV_LOAD_IMAGE_GRAYSCALE)
        assert(img is not None)
        logger("process %s" % imgPath)
        # Paper Table2 jitter F-layer
        f_bbox = bbox.subBBox(-0.05, 1.05, -0.05, 1.05)
        #f_bbox = bbox.subBBox(0, 1, 0, 1)
        f_face = img[f_bbox.top:f_bbox.bottom+1,f_bbox.left:f_bbox.right+1]
        
        
        ## data argument
        if argument and np.random.rand() > -1:
            ### flip
            face_flipped, landmark_flipped = flip(f_face, landmarkGt)
            face_flipped = cv2.resize(face_flipped, (39, 39))
            F_imgs.append(face_flipped.reshape((1, 39, 39)))
            F_landmarks.append(landmark_flipped.reshape(10))
            ### rotation +5 degrees
            if np.random.rand() > 0.5:
                face_rotated_by_alpha, landmark_rotated = rotate(img, f_bbox, \
                    bbox.reprojectLandmark(landmarkGt), 5)
                landmark_rotated = bbox.projectLandmark(landmark_rotated)
                face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (39, 39))
                F_imgs.append(face_rotated_by_alpha.reshape((1, 39, 39)))
                F_landmarks.append(landmark_rotated.reshape(10))
                ### flip with rotation
                face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
                face_flipped = cv2.resize(face_flipped, (39, 39))
                F_imgs.append(face_flipped.reshape((1, 39, 39)))
                F_landmarks.append(landmark_flipped.reshape(10))
            ### rotation -5 degrees
            if np.random.rand() > 0.5:
                face_rotated_by_alpha, landmark_rotated = rotate(img, f_bbox, \
                    bbox.reprojectLandmark(landmarkGt), -5)
                landmark_rotated = bbox.projectLandmark(landmark_rotated)
                face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (39, 39))
                F_imgs.append(face_rotated_by_alpha.reshape((1, 39, 39)))
                F_landmarks.append(landmark_rotated.reshape(10))
                ### flip with rotation
                face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
                face_flipped = cv2.resize(face_flipped, (39, 39))
                F_imgs.append(face_flipped.reshape((1, 39, 39)))
                F_landmarks.append(landmark_flipped.reshape(10))


        f_face = cv2.resize(f_face, (39, 39))
        en_face = f_face[:31, :]
        nm_face = f_face[8:, :]


        f_face = f_face.reshape((1, 39, 39))
        f_landmark = landmarkGt.reshape((10))
        F_imgs.append(f_face)
        F_landmarks.append(f_landmark)
        
        ## data argument for EN   
        if argument and np.random.rand() > 0.5:
            ### flip
            face_flipped, landmark_flipped = flip(en_face, landmarkGt)
            face_flipped = cv2.resize(face_flipped, (31, 39)).reshape((1, 31, 39))
            landmark_flipped = landmark_flipped[:3, :].reshape((6))
            EN_imgs.append(face_flipped)
            EN_landmarks.append(landmark_flipped)


        en_face = cv2.resize(en_face, (31, 39)).reshape((1, 31, 39))
        en_landmark = landmarkGt[:3, :].reshape((6))
        EN_imgs.append(en_face)
        EN_landmarks.append(en_landmark) 
        ## data argument for NM
        
        if argument and np.random.rand() > 0.5:
            ### flip
            face_flipped, landmark_flipped = flip(nm_face, landmarkGt)
            face_flipped = cv2.resize(face_flipped, (31, 39)).reshape((1, 31, 39))
            landmark_flipped = landmark_flipped[2:, :].reshape((6))
            NM_imgs.append(face_flipped)
            NM_landmarks.append(landmark_flipped)


        nm_face = cv2.resize(nm_face, (31, 39)).reshape((1, 31, 39))
        nm_landmark = landmarkGt[2:, :].reshape((6))
        NM_imgs.append(nm_face)
        NM_landmarks.append(nm_landmark)
    #Convert the list to array
    F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
    EN_imgs, EN_landmarks = np.asarray(EN_imgs), np.asarray(EN_landmarks)
    NM_imgs, NM_landmarks = np.asarray(NM_imgs),np.asarray(NM_landmarks)
    ### normalize the data and shu
    F_imgs = processImage(F_imgs)
    shuffle_in_unison_scary(F_imgs, F_landmarks)
    EN_imgs = processImage(EN_imgs)
    shuffle_in_unison_scary(EN_imgs, EN_landmarks)
    NM_imgs = processImage(NM_imgs)
    shuffle_in_unison_scary(NM_imgs, NM_landmarks)
    
    # full face
    base = join(OUTPUT, '1_F')
    createDir(base)
    output = join(base, fname)
    logger("generate %s" % output)
    with h5py.File(output, 'w') as h5:
        h5['data'] = F_imgs.astype(np.float32)
        h5['landmark'] = F_landmarks.astype(np.float32)
        
    # eye and nose
    base = join(OUTPUT, '1_EN')
    createDir(base)
    output = join(base, fname)
    logger("generate %s" % output)
    with h5py.File(output, 'w') as h5:
        h5['data'] = EN_imgs.astype(np.float32)
        h5['landmark'] = EN_landmarks.astype(np.float32)


    # nose and mouth
    base = join(OUTPUT, '1_NM')
    createDir(base)
    output = join(base, fname)
    logger("generate %s" % output)
    with h5py.File(output, 'w') as h5:
        h5['data'] = NM_imgs.astype(np.float32)
        h5['landmark'] = NM_landmarks.astype(np.float32)
if __name__ == '__main__':
    # train data
    train_txt = join(TRAIN, 'trainImageList.txt')
    generate_hdf5_data(train_txt, OUTPUT, 'train.h5', argument=True)


    test_txt = join(TRAIN, 'testImageList.txt')
    generate_hdf5_data(test_txt, OUTPUT, 'test.h5')


    with open(join(OUTPUT, '1_F/train.txt'), 'w') as fd:
        fd.write('train/1_F/train.h5')
    with open(join(OUTPUT, '1_EN/train.txt'), 'w') as fd:
        fd.write('train/1_EN/train.h5')
    with open(join(OUTPUT, '1_NM/train.txt'), 'w') as fd:
        fd.write('train/1_NM/train.h5')
    with open(join(OUTPUT, '1_F/test.txt'), 'w') as fd:
        fd.write('train/1_F/test.h5')
    with open(join(OUTPUT, '1_EN/test.txt'), 'w') as fd:
        fd.write('train/1_EN/test.h5')
    with open(join(OUTPUT, '1_NM/test.txt'), 'w') as fd:
        fd.write('train/1_NM/test.h5')
    # Done

然后得到训练数据，如图：

3、训练，有了数据就可以训练了。老规矩：新建txt然后改为bat文件

.\Build\x64\Release\caffe.exe train --solver=.\try1_2\1_F_solver.prototxt
pause

4、测试。

Build\x64\Release\caffe.exe test --model=try1_2\1_F_train.prototxt -weights=try1_2\_iter_50000.caffemodel
pause

这错误率好的吓人。

5）检测代码

找到了没成功，让我再研究一下。