unet网络搭建代码入门

UNet是一种深度学习网络架构，常用于图像分割任务，尤其是在医学影像分析中。它结合了上采样（upsampling）和下采样（downsampling）结构，以生成具有详细局部信息的特征图。以下是一个简单的Python入门示例，使用Keras库来搭建UNet：

from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, Dropout, UpSampling2D, Concatenate
from tensorflow.keras.models import Model

def conv_block(input_tensor, n_filters, kernel_size=3, activation='relu'):
    x = Conv2D(n_filters, kernel_size, padding='same')(input_tensor)
    x = Activation(activation)(x)
    x = BatchNormalization()(x) # 使用批量归一化加速训练
    return x

def unet(input_shape=(512, 512, 3), n_classes=1):
    inputs = Input(input_shape)

    # 编码部分
    encoder = conv_block(inputs, 64)
    encoder = conv_block(encoder, 128)
    pool1 = MaxPooling2D(pool_size=(2, 2))(encoder)
    encoder = conv_block(pool1, 256)
    pool2 = MaxPooling2D(pool_size=(2, 2))(encoder)
    encoder = conv_block(pool2, 512)

    # 解码部分
    decoder = UpSampling2D(size=(2, 2))(encoder)
    decoder = Concatenate()([decoder, conv_block(pool2, 512, kernel_size=1)])
    decoder = conv_block(decoder, 512)
    decoder = UpSampling2D(size=(2, 2))(decoder)
    decoder = Concatenate()([decoder, conv_block(pool1, 256, kernel_size=1)])
    decoder = conv_block(decoder, 256)

    # 输出层
    output = Conv2D(n_classes, 1, activation='softmax')(decoder)

    model = Model(inputs=inputs, outputs=output)
    return model

# 创建并编译模型
model = unet()
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

在这个例子中，`unet`函数返回了一个完整的UNet模型，你可以调整输入形状、滤波器数量等参数以适应你的需求。