1.VGG
VGG全称是Visual Geometry Group属于牛津大学科学工程系，其发布了一些列以VGG开头的卷积网络模型，可以应用在人脸识别、图像分类等方面,VGG的输入被设置为大小为224x244的RGB图像。为训练集图像上的所有图像计算平均RGB值，然后将该图像作为输入输入到VGG卷积网络。使用3x3或1x1滤波器，并且卷积步骤是固定的。有3个VGG全连接层，根据卷积层+全连接层的总数，可以从VGG11到VGG19变化。最小VGG11具有8个卷积层和3个完全连接层。最大VGG19具有16个卷积层+3个完全连接的层。此外，VGG网络后面没有每个卷积层后面的池化层，也没有分布在不同卷积层下的总共5个池化层。
结构图如下：
架构图
2.Unet模型：
Unet是一个优秀的语义分割模型，其主要执行过程与其它语义分割模型类似。与CNN不同的之处在于CNN是图像级的分类，而unet是像素级的分类，其输出的是每个像素点的类别
主要代码如下：
def get_vgg_encoder(input_height=224, input_width=224, channels=3):
if channel == 'channels_first':
img_input = Input(shape=(channels, input_height, input_width))
elif channel == 'channels_last':
img_input = Input(shape=(input_height, input_width, channels))
x = Conv2D(64, (3, 3), activation='relu', padding='same',
name='block1_conv1', data_format=channel)(img_input)
x = Conv2D(64, (3, 3), activation='relu', padding='same',
name='block1_conv2', data_format=channel)(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool',
data_format=channel)(x)
f1 = x
# Block 2
x = Conv2D(128, (3, 3), activation='relu', padding='same',
name='block2_conv1', data_format=channel)(x)
x = Conv2D(128, (3, 3), activation='relu', padding='same',
name='block2_conv2', data_format=channel)(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool',
data_format=channel)(x)
f2 = x
# Block 3
x = Conv2D(256, (3, 3), activation='relu', padding='same',
name='block3_conv1', data_format=channel)(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same',
name='block3_conv2', data_format=channel)(x)
x = Conv2D(256, (3, 3), activation='relu', padding='same',
name='block3_conv3', data_format=channel)(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool',
data_format=channel)(x)
f3 = x
# Block 4
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block4_conv1', data_format=channel)(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block4_conv2', data_format=channel)(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block4_conv3', data_format=channel)(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool',
data_format=channel)(x)
f4 = x
# Block 5
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block5_conv1', data_format=channel)(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block5_conv2', data_format=channel)(x)
x = Conv2D(512, (3, 3), activation='relu', padding='same',
name='block5_conv3', data_format=channel)(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool',
data_format=channel)(x)
f5 = x
return img_input, [f1, f2, f3, f4, f5]
def _unet(classes, encoder, l1_skip_conn=True, input_height=416,
input_width=608, channels=3):
img_input, levels = encoder(
input_height=input_height, input_width=input_width, channels=channels)
[f1, f2, f3, f4, f5] = levels
o = f4
o = (ZeroPadding2D((1, 1), data_format=channel))(o)
o = (Conv2D(512, (3, 3), padding='valid' , activation='relu' , data_format=channel))(o)
o = (BatchNormalization())(o)
o = (UpSampling2D((2, 2), data_format=channel))(o)
o = (concatenate([o, f3], axis=-1))
o = (ZeroPadding2D((1, 1), data_format=channel))(o)
o = (Conv2D(256, (3, 3), padding='valid', activation='relu' , data_format=channel))(o)
o = (BatchNormalization())(o)
o = (UpSampling2D((2, 2), data_format=channel))(o)
o = (concatenate([o, f2], axis=-1))
o = (ZeroPadding2D((1, 1), data_format=channel))(o)
o = (Conv2D(128, (3, 3), padding='valid' , activation='relu' , data_format=channel))(o)
o = (BatchNormalization())(o)
o = (UpSampling2D((2, 2), data_format=channel))(o)
if l1_skip_conn:
o = (concatenate([o, f1], axis=-1))
o = (ZeroPadding2D((1, 1), data_format=channel))(o)
o = (Conv2D(64, (3, 3), padding='valid', activation='relu', data_format=channel, name="seg_feats"))(o)
o = (BatchNormalization())(o)
o = Conv2D(classes, (3, 3), padding='same',
data_format=channel)(o)
model = get_segmentation_model(img_input, o)
return model
