主页:https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html 论文:https://arxiv.org/abs/1604.02878 代码:官方matlab版、C++ caffe版 第三方训练代码:tensorflow、mxnet
MTCNN,恰如论文标题《Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks》所言,采用级联CNN结构,通过多任务学习,同时完成了两个任务——人脸检测和人脸对齐,输出人脸的Bounding Box以及人脸的关键点(眼睛、鼻子、嘴)位置。
MTCNN 又好又快,提出时在FDDB、WIDER FACE和AFLW数据集上取得了当时(2016年4月)最好的结果,速度又快,现在仍被广泛使用作为人脸识别的前端,如InsightFace和facenet。
MTCNN效果为什么好,文中提了3个主要的原因:
下面详细介绍。
总体而言,MTCNN方法可以概括为:图像金字塔+3阶段级联CNN,如下图所示
对输入图像建立金字塔是为了检测不同尺度的人脸,通过级联CNN完成对人脸 由粗到细(coarse-to-fine) 的检测,所谓级联指的是 前者的输出是后者的输入,前者往往先使用少量信息做个大致的判断,快速将不是人脸的区域剔除,剩下可能包含人脸的区域交给后面更复杂的网络,利用更多信息进一步筛选,这种由粗到细的方式在保证召回率的情况下可以大大提高筛选效率。下面为MTCNN中级联的3个网络(P-Net、R-Net、O-Net),可以看到它们的网络层数逐渐加深,输入图像的尺寸(感受野)在逐渐变大12→24→48,最终输出的特征维数也在增加32→128→256,意味着利用的信息越来越多。
工作流程是怎样的? 首先,对原图通过双线性插值构建图像金字塔,可以参看前面的博文《人脸检测中,如何构建输入图像金字塔》。构建好金字塔后,将金字塔中的图像逐个输入给P-Net。
需要注意的是:
MTCNN效果好的第1个原因是精心设计的级联CNN架构,其实,级联的思想早已有之,而使用级联CNN进行人脸检测的方法是在2015 CVPR《A convolutional neural network cascade for face detection》中被率先提出,MTCNN与之的差异在于:
这样使网络的表达能力更强,同时运行时间更少。
MTCNN效果好的后面2个原因在线困难样本挖掘和人脸对齐联合学习将在下一节介绍。
MTCNN的多任务学习有3个任务,1个分类2个回归,分别为face classification、bounding box regression以及facial landmark localization,分类的损失函数使用交叉熵损失,回归的损失函数使用欧氏距离损失,如下:
MTCNN准备了4种训练数据:
lable = 0
lable = 1
lable = -1
lable = -2
这4种数据是如何组织的呢?以MTCNN-Tensorflow为例:
Since MTCNN is a Multi-task Network,we should pay attention to the format of training data.The format is: [path to image] [cls_label] [bbox_label] [landmark_label] For neg sample, cls_label=0, bbox_label=[0,0,0,0], landmark_label=[0,0,0,0,0,0,0,0,0,0]. For pos sample, cls_label=1, bbox_label(calculate), landmark_label=[0,0,0,0,0,0,0,0,0,0]. For part sample, cls_label=-1, bbox_label(calculate), landmark_label=[0,0,0,0,0,0,0,0,0,0]. For landmark sample, cls_label=-2, bbox_label=[0,0,0,0], landmark_label(calculate).
数量之比依次为\(3:1:1:2\),其中,Negatives、Positives和Part faces通过WIDER FACE数据集crop得到,landmark faces通过CelebA数据集crop得到,先crop区域,然后看这个区域与哪个ground-truth face的IOU最大,根据最大IOU来生成label,比如小于0.3的标记为negative。
P-Net训练数据的准备可以参见gen_12net_data.py、gen_landmark_aug_12.py、gen_imglist_pnet.py和gen_PNet_tfrecords.py,代码很直观,这里略过crop过程,重点介绍bounding box label和landmark label的生成。下面是gen_12net_data.py和gen_landmark_aug_12.py中的代码片段,bounding box 和 landmark 的label为归一化后的相对坐标,offset_x1, offset_y1, offset_x2, offset_y2
为bounding box的label,使用crop区域的size进行归一化,rv
为landmark的label,使用bbox的宽高进行归一化,注意两者的归一化是不一样的,具体见代码:
## in gen_12net_data.py
# pos and part face size [minsize*0.8,maxsize*1.25]
size = npr.randint(int(min(w, h) * 0.8), np.ceil(1.25 * max(w, h)))
# delta here is the offset of box center
if w<5:
print (w)
continue
#print (box)
delta_x = npr.randint(-w * 0.2, w * 0.2)
delta_y = npr.randint(-h * 0.2, h * 0.2)
#show this way: nx1 = max(x1+w/2-size/2+delta_x)
# x1+ w/2 is the central point, then add offset , then deduct size/2
# deduct size/2 to make sure that the right bottom corner will be out of
nx1 = int(max(x1 + w / 2 + delta_x - size / 2, 0))
#show this way: ny1 = max(y1+h/2-size/2+delta_y)
ny1 = int(max(y1 + h / 2 + delta_y - size / 2, 0))
nx2 = nx1 + size
ny2 = ny1 + size
if nx2 > width or ny2 > height:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
#yu gt de offset
##### x1 y1 x2 y2 为 ground truth bbox, nx1 ny1 nx2 ny2为crop的区域,size为crop的区域size ######
offset_x1 = (x1 - nx1) / float(size)
offset_y1 = (y1 - ny1) / float(size)
offset_x2 = (x2 - nx2) / float(size)
offset_y2 = (y2 - ny2) / float(size)
#crop
cropped_im = img[ny1 : ny2, nx1 : nx2, :]
#resize
resized_im = cv2.resize(cropped_im, (12, 12), interpolation=cv2.INTER_LINEAR)
##########################################################################
## in gen_landmark_aug_12.py
#normalize land mark by dividing the width and height of the ground truth bounding box
# landmakrGt is a list of tuples
for index, one in enumerate(landmarkGt):
# (( x - bbox.left)/ width of bounding box, (y - bbox.top)/ height of bounding box
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1]))
# put the normalized value into the new list landmark
landmark[index] = rv
需要注意的是,对于P-Net,其为FCN,预测阶段输入图像可以为任意大小,但在训练阶段,使用的训练数据均被resize到12×12,以便于控制正负样本的比例(避免数据不平衡)。
因为是级联结构,训练要分阶段依次进行,训练好P-Net后,用P-Net产生的候选区域来训练R-Net,训练好R-Net后,再生成训练数据来训练O-Net。P-Net训练好之后,根据其结果准备R-Net的训练数据,R-Net训练好之后,再准备O-Net的训练数据,过程是类似的,具体可以参见相关代码,这里就不赘述了。
4种训练数据参与的训练任务如下:
至于在线困难样本挖掘,仅在训练face/non-face classification时使用,具体做法是:对每个mini-batch的数据先通过前向传播,挑选损失最大的前70%作为困难样本,在反向传播时仅使用这70%困难样本产生的损失。文中的实验表明,这样做在FDDB数据级上可以带来1.5个点的性能提升。
具体怎么实现的?这里以MTCNN-Tensorflow / train_models / mtcnn_model.py代码为例,用label
来指示是哪种数据,下面为代码,重点关注valid_inds
和loss
(square_error
)的计算(对应\(\beta_i^j\)),以及cls_ohem
中的困难样本挖掘。
# in mtcnn_model.py]
# pos=1, neg=0, part=-1, landmark=-2
# 通过cls_ohem, bbox_ohem, landmark_ohem来计算损失
num_keep_radio = 0.7 # mini-batch前70%做为困难样本
# face/non-face 损失,注意在线困难样本挖掘(前70%)
def cls_ohem(cls_prob, label):
zeros = tf.zeros_like(label)
#label=-1 --> label=0net_factory
#pos -> 1, neg -> 0, others -> 0
label_filter_invalid = tf.where(tf.less(label,0), zeros, label)
num_cls_prob = tf.size(cls_prob)
cls_prob_reshape = tf.reshape(cls_prob,[num_cls_prob,-1])
label_int = tf.cast(label_filter_invalid,tf.int32)
# get the number of rows of class_prob
num_row = tf.to_int32(cls_prob.get_shape()[0])
#row = [0,2,4.....]
row = tf.range(num_row)*2
indices_ = row + label_int
label_prob = tf.squeeze(tf.gather(cls_prob_reshape, indices_))
loss = -tf.log(label_prob+1e-10)
zeros = tf.zeros_like(label_prob, dtype=tf.float32)
ones = tf.ones_like(label_prob,dtype=tf.float32)
# set pos and neg to be 1, rest to be 0
valid_inds = tf.where(label < zeros,zeros,ones)
# get the number of POS and NEG examples
num_valid = tf.reduce_sum(valid_inds)
###### 困难样本数量 #####
keep_num = tf.cast(num_valid*num_keep_radio,dtype=tf.int32)
#FILTER OUT PART AND LANDMARK DATA
loss = loss * valid_inds
loss,_ = tf.nn.top_k(loss, k=keep_num) ##### 仅取困难样本反向传播 #####
return tf.reduce_mean(loss)
# bounding box损失
#label=1 or label=-1 then do regression
def bbox_ohem(bbox_pred,bbox_target,label):
'''
:param bbox_pred:
:param bbox_target:
:param label: class label
:return: mean euclidean loss for all the pos and part examples
'''
zeros_index = tf.zeros_like(label, dtype=tf.float32)
ones_index = tf.ones_like(label,dtype=tf.float32)
# keep pos and part examples
valid_inds = tf.where(tf.equal(tf.abs(label), 1),ones_index,zeros_index)
#(batch,)
#calculate square sum
square_error = tf.square(bbox_pred-bbox_target)
square_error = tf.reduce_sum(square_error,axis=1)
#keep_num scalar
num_valid = tf.reduce_sum(valid_inds)
#keep_num = tf.cast(num_valid*num_keep_radio,dtype=tf.int32)
# count the number of pos and part examples
keep_num = tf.cast(num_valid, dtype=tf.int32)
#keep valid index square_error
square_error = square_error*valid_inds
# keep top k examples, k equals to the number of positive examples
_, k_index = tf.nn.top_k(square_error, k=keep_num)
square_error = tf.gather(square_error, k_index)
return tf.reduce_mean(square_error)
# 关键点损失
def landmark_ohem(landmark_pred,landmark_target,label):
'''
:param landmark_pred:
:param landmark_target:
:param label:
:return: mean euclidean loss
'''
#keep label =-2 then do landmark detection
ones = tf.ones_like(label,dtype=tf.float32)
zeros = tf.zeros_like(label,dtype=tf.float32)
valid_inds = tf.where(tf.equal(label,-2),ones,zeros) ##### 将label=-2的置为1,其余为0 #####
square_error = tf.square(landmark_pred-landmark_target)
square_error = tf.reduce_sum(square_error,axis=1)
num_valid = tf.reduce_sum(valid_inds)
#keep_num = tf.cast(num_valid*num_keep_radio,dtype=tf.int32)
keep_num = tf.cast(num_valid, dtype=tf.int32)
square_error = square_error*valid_inds # 在计算landmark_ohem损失时只计算beta=1的 #####
_, k_index = tf.nn.top_k(square_error, k=keep_num)
square_error = tf.gather(square_error, k_index)
return tf.reduce_mean(square_error)
多任务学习的代码片段如下:
# in train.py
if net == 'PNet':
image_size = 12
radio_cls_loss = 1.0;radio_bbox_loss = 0.5;radio_landmark_loss = 0.5;
elif net == 'RNet':
image_size = 24
radio_cls_loss = 1.0;radio_bbox_loss = 0.5;radio_landmark_loss = 0.5;
else:
radio_cls_loss = 1.0;radio_bbox_loss = 0.5;radio_landmark_loss = 1;
image_size = 48
# ...
# 多任务联合损失
total_loss_op = radio_cls_loss*cls_loss_op + radio_bbox_loss*bbox_loss_op + radio_landmark_loss*landmark_loss_op + L2_loss_op
train_op, lr_op = train_model(base_lr, total_loss_op, num)
def train_model(base_lr, loss, data_num):
"""
train model
:param base_lr: base learning rate
:param loss: loss
:param data_num:
:return:
train_op, lr_op
"""
lr_factor = 0.1
global_step = tf.Variable(0, trainable=False)
#LR_EPOCH [8,14]
#boundaried [num_batch,num_batch]
boundaries = [int(epoch * data_num / config.BATCH_SIZE) for epoch in config.LR_EPOCH]
#lr_values[0.01,0.001,0.0001,0.00001]
lr_values = [base_lr * (lr_factor ** x) for x in range(0, len(config.LR_EPOCH) + 1)]
#control learning rate
lr_op = tf.train.piecewise_constant(global_step, boundaries, lr_values)
optimizer = tf.train.MomentumOptimizer(lr_op, 0.9)
train_op = optimizer.minimize(loss, global_step)
return train_op, lr_op
以上对应论文中的损失函数。
而每个位置处都有个4维的向量,其为bounding box左上角和右下角的偏移dx1, dy1, dx2, dy2
,通过上面的训练过程,我们知道它们是归一化之后的相对坐标,通过对应的区域以及归一化后的相对坐标就可以获得原图上的bounding box,如下所示,dx1, dy1, dx2, dy2
为归一化的相对坐标,求到原图中的bounding box坐标的过程为生成训练数据bounding box label的逆过程。
landmark位置通过O-Net输出得到,将人脸候选框resize到48×48输入给O-Net,先获得bounding box(同上),因为O-Net输出的landmark也是归一化后的相对坐标,通过bounding box的长宽和bounding box左上角求取landmark 在原图中的位置,如下所示:
至此,预测过程中的要点也介绍完毕了,以上。