Tensorflow2.0—FaceNet⽹络原理及代码解析(⼀)---模型原理及⾻⼲⽹络Tensorflow2.0—FaceNet⽹络原理及代码解析(⼀)— 模型原理及⾻⼲⽹络五颗豌豆
公务用车管理规定
乒乓比赛FaceNet其实就是⼀个前⾔所诉的通⽤⼈脸识别系统:采⽤深度卷积神经⽹络(CNN)学习将图像映射到欧式空间。空间距离直接和图⽚相似度相关:同⼀个⼈的不同图像在空间距离很⼩,不同⼈的图像在空间中有较⼤的距离,可以⽤于⼈脸验证、识别和聚类。在800万⼈,2亿多张样本集训练后,FaceNet在LFW数据集上测试的准确率达到了99.63%,在YouTube Faces DB数据集上,准确率为95.12%。
代码:
⼀、算法原理
上图是facenet的⼤体框架⽰意图,可以看到,⼀个batch的训练图⽚经过⼀个⿊盒⼦进⾏特征提取。其中,DEEP ARCHITECTURE这个部分⼀般是⼀个⽐较成熟的backbone,最早的FaceNet采⽤两种深度卷积⽹络:经典Zeiler&Fergus架构和Google的Inception v1。最新的FaceNet进⾏了改进,主体模型采⽤⼀个极深度⽹络Inception ResNet -v2,由3个带有残差连接的Inception模块和1个Inception v4模块组成。在代码实现中通常使⽤mobilenet或者inception_resnetv1作为⾻⼲⽹络。
经过了特征提取之后进⾏L2标准化,然后得到⼀个128维的向量。
最后,进⾏Triplet loss计算。
下⾯,按照每个模块进⾏介绍~
⼆、⾻⼲⽹络
在这⾥,我们选取mobilenet作为⾻⼲⽹络。
在代码中,backbone = "mobilenet",然后代⼊model = facenet(input_shape, num_class, backbone=backbone, mode="train"),mobilenet实现的代码全部在nets/mobilenet.py中。
def_conv_block(inputs, filters, kernel=(3,3), strides=(1,1)):
x = Conv2D(filters, kernel,
padding='same',
u_bias=Fal,
strides=strides,
name='conv1')(inputs)
x = BatchNormalization(name='conv1_bn')(x)
return Activation(relu6, name='conv1_relu')(x)
u盘检测不到这是典型的卷积-BN-激活模块,没啥好说的~
def_depthwi_conv_block(inputs, pointwi_conv_filters, depth_multiplier=1, strides=(1,1), block_id=1):
x = DepthwiConv2D((3,3),
padding='same',
depth_multiplier=depth_multiplier,
strides=strides,
u_bias=Fal,
name='conv_dw_%d'% block_id)(inputs)
x = BatchNormalization(name='conv_dw_%d_bn'% block_id)(x)
x = Activation(relu6, name='conv_dw_%d_relu'% block_id)(x)
x = Conv2D(pointwi_conv_filters,(1,1),
padding='same',
u_bias=Fal,
strides=(1,1),
name='conv_pw_%d'% block_id)(x)ppt页码怎么设置
x = BatchNormalization(name='conv_pw_%d_bn'% block_id)(x)
return Activation(relu6, name='conv_pw_%d_relu'% block_id)(x)
def MobileNet(inputs, embedding_size=128, dropout_keep_prob=0.4, alpha=1.0, depth_multiplier=1):
# 160,160,3 -> 80,80,32
x = _conv_block(inputs,32, strides=(2,2))
鸭肠的做法
# 80,80,32 -> 80,80,64
x = _depthwi_conv_block(x,64, depth_multiplier, block_id=1)
# 80,80,64 -> 40,40,128
x = _depthwi_conv_block(x,128, depth_multiplier, strides=(2,2), block_id=2)
x = _depthwi_conv_block(x,128, depth_multiplier, block_id=3)
# 40,40,128 -> 20,20,256
x = _depthwi_conv_block(x,256, depth_multiplier, strides=(2,2), block_id=4)
x = _depthwi_conv_block(x,256, depth_multiplier, block_id=5)
# 20,20,256 -> 10,10,512
x = _depthwi_conv_block(x,512, depth_multiplier, strides=(2,2), block_id=6)小动物手抄报
x = _depthwi_conv_block(x,512, depth_multiplier, block_id=7)
x = _depthwi_conv_block(x,512, depth_multiplier, block_id=8)
x = _depthwi_conv_block(x,512, depth_multiplier, block_id=9)
x = _depthwi_conv_block(x,512, depth_multiplier, block_id=10)
x = _depthwi_conv_block(x,512, depth_multiplier, block_id=11)
# 10,10,512 -> 5,5,1024
x = _depthwi_conv_block(x,1024, depth_multiplier, strides=(2,2), block_id=12)
x = _depthwi_conv_block(x,1024, depth_multiplier, block_id=13)
# 1024
x = GlobalAveragePooling2D()(x)
# 防⽌⽹络过拟合,训练的时候起作⽤
诛仙之白衣凌波x = Dropout(1.0- dropout_keep_prob, name='Dropout')(x)
# 全连接层到128
# 128
x = Den(embedding_size, u_bias=Fal, name='Bottleneck')(x)
x = BatchNormalization(momentum=0.995, epsilon=0.001, scale=Fal,
name='BatchNorm_Bottleneck')(x)
# 创建模型
model = Model(inputs, x, name='mobilenet')
return model
上述就是mobilenet的⽹络结构,看得出来,实现起来很简洁明了,(160,160,3)的输⼊图⽚经过了DEEP ARCHITECTURE之后⽣成了⼀个(5,5,1024)的feature map,然后进⾏了⼀次GlobalAveragePooling2D(全局平均池化),在这个blog我做了简单的GAP的介绍。GAP层将(5,5,1024)变成了(1,1,1024),再做⼀次Dropout,最后连接⼀个全连接层,让其变为长度是128的向量。
if mode =="train":
#-----------------------------------------#
# 训练的话利⽤交叉熵和triplet_loss
# 结合⼀起训练
#-----------------------------------------#
logits = Den(num_class)(model.output)#(None,10575)
softmax = Activation("softmax", name ="Softmax")(logits)#(None,10575)
normalize = Lambda(lambda x: K.l2_normalize(x, axis=1), name="Embedding")(model.output)#(None,128)
combine_model = Model(inputs,[softmax, normalize])
return combine_model
最后,将经过backbone⽣成的(None,128)向量连接⼀个全连接层和⼀个softmax激活。另外,将(None,128)进⾏L2标准化操作。