def create_context_module(input_layer,
n_level_filters,
dropout_rate=0.3,
data_format="channels_last"):
convolution1 = create_convolution_block(input_layer=input_layer,
n_filters=n_level_filters)
dropout = SpatialDropout2D(rate=dropout_rate,
data_format=data_format)(convolution1)
convolution2 = create_convolution_block(input_layer=dropout,
n_filters=n_level_filters)
return convolution2
def uent2d_model(input_shape=(256, 256, 1),
n_base_fileters=16,
depth=5,
dropout_rate=0.3,
n_segmentation_levels=3,
n_labels=1,
optimizer=Adam,
initial_learning_rate=5e-4,
loss_function=weighted_dice_coefficient_loss,
activation_name="sigmoid"):
"""
:param input_shape:
:param n_base_fileters: 第一个卷积块的滤波器个数
:param depth: unet结构的深度
:param dropout_rate:
:param n_segmentation_levels:
:param n_labels:
:param optimizer:
:param initial_learning_rate:
:param loss_function:
:param activation_name:
:return:
"""
"原始的unet"
inputs = Input(input_shape)
current_layer = inputs
level_output_layers = list()
level_filters = list() # 每个卷积块的滤波器个数
for level_num in range(depth): # 遍历深度
# 计算得到各卷积块的滤波器个数
n_level_filters = (2**level_num) * n_base_fileters
level_filters.append(n_level_filters)
if current_layer is inputs:
in_conv = create_convolution_block(current_layer, n_level_filters)
else:
in_conv = create_convolution_block(current_layer,
n_level_filters,
strides=(2, 2))
# 创建本层的卷积块
context_output_layer = create_context_module(in_conv,
n_level_filters,
dropout_rate=dropout_rate)
# 创建卷积块的残差并且加入到输出列表,多尺度融合采用的方式是相加而不是concatenate能节约内存
summation_layer = Add()([in_conv,
context_output_layer]) # 这里实质上是实现本次的残差块
level_output_layers.append(summation_layer)
current_layer = summation_layer
# 下面进行上采样部分,从最下层开始：current_layer
segmentation_layers = list()
for level_num in range(depth - 2, -1, -1): # 3, 2, 1, 0
up_sampling = create_up_sampling_module(
current_layer, level_filters[level_num]) # 先上采样，再卷积一次
concatenation_layer = Concatenate(axis=-1)(
[level_output_layers[level_num],
up_sampling]) # 当前水平层的通道上的skip connect
localization_output = create_localization_module(
concatenation_layer,
level_filters[level_num]) # 对skip connect之后的结果进行两次卷积操作
current_layer = localization_output
if level_num < n_segmentation_levels:
segmentation_layers.insert(0,
Conv2D(n_labels, (1, 1))(current_layer))
output_layer = None
for level_num in reversed(
range(n_segmentation_levels)): # 0, 1, 2--->2, 1, 0
segmentation_layer = segmentation_layers[level_num]
if output_layer is None:
output_layer = segmentation_layer
else:
output_layer = Add()([output_layer, segmentation_layer])
if level_num > 0: # 2, 1
output_layer = UpSampling2D(size=(2, 2))(output_layer)
activation_block = Activation(activation_name)(output_layer)
model = Model(inputs=inputs, outputs=activation_block)
model.compile(optimizer=optimizer(lr=initial_learning_rate),
loss=loss_function)
return model
def create_localization_module(input_layer, n_filters):
convolution1 = create_convolution_block(input_layer, n_filters)
convolution2 = create_convolution_block(convolution1,
n_filters,
kernel=(1, 1))
return convolution2
def create_up_sampling_module(input_layer, n_filters, size=(2, 2)):
up_sample = UpSampling2D(size=size)(input_layer)
convolution = create_convolution_block(up_sample, n_filters)
return convolution
def uent2d_dilate_model_1X1(input_shape=(256, 256, 1),
n_base_fileters=16,
depth=5,
dropout_rate=0.3,
n_segmentation_levels=3,
n_labels=1,
optimizer=Adam,
initial_learning_rate=5e-4,
loss_function=weighted_dice_coefficient_loss,
activation_name="sigmoid"):
"""
:param input_shape:
:param n_base_fileters: 第一个卷积块的滤波器个数
:param depth: unet结构的深度
:param dropout_rate:
:param n_segmentation_levels:
:param n_labels:
:param optimizer:
:param initial_learning_rate:
:param loss_function:
:param activation_name:
:return:
"""
# 在buttonblock添加三个并联的空洞卷积率分别为2， 3， 5的分支，结果和buttonblock的输出执行concat操作
# 执行完concat操作之后，使用多层1*1*C的卷积核进行特征融合增加非线性，
# 注意，此处三个并联的空洞卷积的输出可以尝试分别使用sum和concat
# 先concat之后再使用1*1卷积核效果会好一些？
# 直接将4个分支进行concat，然后使用1*1卷积融合
dilate_rate = [2, 3, 5]
inputs = Input(input_shape)
current_layer = inputs
level_output_layers = list()
level_filters = list() # 每个卷积块的滤波器个数
for level_num in range(depth): # 遍历深度
# 计算得到各卷积块的滤波器个数
n_level_filters = (2**level_num) * n_base_fileters
level_filters.append(n_level_filters)
if current_layer is inputs:
in_conv = create_convolution_block(current_layer, n_level_filters)
else:
in_conv = create_convolution_block(current_layer,
n_level_filters,
strides=(2, 2))
# 创建本层的卷积块
context_output_layer = create_context_module(in_conv,
n_level_filters,
dropout_rate=dropout_rate)
# 创建卷积块的残差并且加入到输出列表,多尺度融合采用的方式是相加而不是concatenate能节约内存
summation_layer = Add()([in_conv,
context_output_layer]) # 这里实质上是实现本次的残差块
level_output_layers.append(summation_layer)
current_layer = summation_layer
# 在此处添加空洞卷积操作，即在buttonblock处添加dilation
filters = level_filters[-1]
dilate_features = list()
dilate_features.append(current_layer)
for rate in dilate_rate:
dilate_feature = create_dilate_convolution_block(current_layer,
filters,
dilate_rate=rate)
feature = create_localization_module(dilate_feature, filters)
dilate_features.append(feature)
# current_layer = Add()(dilate_features)
merged_feature = Concatenate()(dilate_features) # (?, 32, 32, 1024)
current_layer = create_1x1_convolution_module(merged_feature, filters)
# 下面进行上采样部分,从最下层开始：current_layer
segmentation_layers = list()
for level_num in range(depth - 2, -1, -1): # 3, 2, 1, 0
up_sampling = create_up_sampling_module(current_layer,
level_filters[level_num])
# 先上采样，再卷积一次
concatenation_layer = Concatenate(axis=-1)(
[level_output_layers[level_num], up_sampling])
# 当前水平层的通道上的skip connect
localization_output = create_localization_module(
concatenation_layer, level_filters[level_num])
# 对skip connect之后的结果进行两次卷积操作
current_layer = localization_output
if level_num < n_segmentation_levels:
segmentation_layers.insert(0,
Conv2D(n_labels, (1, 1))(current_layer))
output_layer = None
for level_num in reversed(
range(n_segmentation_levels)): # 0, 1, 2--->2, 1, 0
segmentation_layer = segmentation_layers[level_num]
if output_layer is None:
output_layer = segmentation_layer
else:
output_layer = Add()([output_layer, segmentation_layer])
if level_num > 0: # 2, 1
output_layer = UpSampling2D(size=(2, 2))(output_layer)
activation_block = Activation(activation_name)(output_layer)
model = Model(inputs=inputs, outputs=activation_block)
model.compile(optimizer=optimizer(lr=initial_learning_rate),
loss=loss_function)
return model