keras实现U-Net, R2U-Net, Attention U-Net, Attention R2U-Net

keras实现U-Net, R2U-Net, Attention U-Net, Attention R2U-Net
代码在github上，，记得给星哦。
我的github地址
https://github.com/lixiaolei1982/Keras-Implementation-of-U-Net-R2U-Net-Attention-U-Net-Attention-R2U-Net.-
一，Unet结构：

优点：结构简单易懂，能在很小的数据集训练并取得不错的效果，用于许多的生物医学的分割。
缺点：网络深度不够，导致在多分类中表现一般。
参考代码：

def unet(img_w, img_h, n_label, data_format='channels_first'):
    inputs = Input((3, img_w, img_h))
    x = inputs
    depth = 4
    features = 64
    skips = []
    for i in range(depth):
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        x = Dropout(0.2)(x)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        skips.append(x)
        x = MaxPooling2D((2, 2), data_format= data_format)(x)
        features = features * 2

    x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
    x = Dropout(0.2)(x)
    x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)

    for i in reversed(range(depth)):
        features = features // 2
        # attention_up_and_concate(x,[skips[i])
        x = UpSampling2D(size=(2, 2), data_format=data_format)(x)
        x = concatenate([skips[i], x], axis=1)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        x = Dropout(0.2)(x)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)

    conv6 = Conv2D(n_label, (1, 1), padding='same', data_format=data_format)(x)
    conv7 = core.Activation('sigmoid')(conv6)
    model = Model(inputs=inputs, outputs=conv7)

    #model.compile(optimizer=Adam(lr=1e-5), loss=[focal_loss()], metrics=['accuracy', dice_coef])
    return model

二， R2U-Net
Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation。

优点：在Unet基础上增加残差与循环网络，网络加深又避免梯度无法学习，用于医学影像的分割。分割效果好于unet。
缺点：model大小及参数远超unet。

def r2_unet(img_w, img_h, n_label, data_format='channels_first'):
    inputs = Input((3, img_w, img_h))
    x = inputs
    depth = 4
    features = 64
    skips = []
    for i in range(depth):
        x = rec_res_block(x, features, data_format=data_format)
        skips.append(x)
        x = MaxPooling2D((2, 2), data_format=data_format)(x)

        features = features * 2

    x = rec_res_block(x, features, data_format=data_format)

    for i in reversed(range(depth)):
        features = features // 2
        x = up_and_concate(x, skips[i], data_format=data_format)
        x = rec_res_block(x, features, data_format=data_format)

    conv6 = Conv2D(n_label, (1, 1), padding='same', data_format=data_format)(x)
    conv7 = core.Activation('sigmoid')(conv6)
    model = Model(inputs=inputs, outputs=conv7)
    #model.compile(optimizer=Adam(lr=1e-6), loss=[dice_coef_loss], metrics=['accuracy', dice_coef])
    return model

三， Att-Net
Attention U-Net: Learning Where to Look for the Pancreas

优点：在Unet基础上增加注意门，使跳远链接更有选择性。
缺点：网络未加深。

def att_unet(img_w, img_h, n_label, data_format='channels_first'):
    inputs = Input((3, img_w, img_h))
    x = inputs
    depth = 4
    features = 64
    skips = []
    for i in range(depth):
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        x = Dropout(0.2)(x)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        skips.append(x)
        x = MaxPooling2D((2, 2), data_format='channels_first')(x)
        features = features * 2

    x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
    x = Dropout(0.2)(x)
    x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)

    for i in reversed(range(depth)):
        features = features // 2
        x = attention_up_and_concate(x, skips[i], data_format=data_format)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)
        x = Dropout(0.2)(x)
        x = Conv2D(features, (3, 3), activation='relu', padding='same', data_format=data_format)(x)

    conv6 = Conv2D(n_label, (1, 1), padding='same', data_format=data_format)(x)
    conv7 = core.Activation('sigmoid')(conv6)
    model = Model(inputs=inputs, outputs=conv7)

    #model.compile(optimizer=Adam(lr=1e-5), loss=[focal_loss()], metrics=['accuracy', dice_coef])
    return model

四，Attention R2U-Net : integration of two recent advanced works (R2U-Net + Attention U-Net)

优点：在Unet基础上增加残差与循环网络，网络加深又避免梯度无法学习，增加注意门，学习更专注有用的东西。
缺点：model大小及参数远超unet。

def att_r2_unet(img_w, img_h, n_label, data_format='channels_first'):
    inputs = Input((3, img_w, img_h))
    x = inputs
    depth = 4
    features = 64
    skips = []
    for i in range(depth):
        x = rec_res_block(x, features, data_format=data_format)
        skips.append(x)
        x = MaxPooling2D((2, 2), data_format=data_format)(x)

        features = features * 2

    x = rec_res_block(x, features, data_format=data_format)

    for i in reversed(range(depth)):
        features = features // 2
        x = attention_up_and_concate(x, skips[i], data_format=data_format)
        x = rec_res_block(x, features, data_format=data_format)

    conv6 = Conv2D(n_label, (1, 1), padding='same', data_format=data_format)(x)
    conv7 = core.Activation('sigmoid')(conv6)
    model = Model(inputs=inputs, outputs=conv7)
    #model.compile(optimizer=Adam(lr=1e-6), loss=[dice_coef_loss], metrics=['accuracy', dice_coef])
    return model

五，Attention R2U-Net 实现遥感卫星图的建筑物分割效果