Awesome
unet
unet主要用于语义分割, 这里是一个细胞边缘检测的例子, 数据集比较简单。 unet的网络结构, 因像字母‘U’而得名。
这里有一篇关于unet的 [论文](U-Net: Convolutional Networks for Biomedical Image Segmentation), 论文里面的网络结构如下:
说一下这个网络: 输入572×572×1, 输出:388×388×2, 大小不一样。 主要是因为卷积的过程中, 每次卷积会减小, 在copy and crop中, 也会减小。
我这里设计的网络, 并没有像上图的网络一样, 原封不动的实现出来, 而是借助vgg网络结构来实现的。
看上图, 我们发现, unet的前半部分采用2层卷积+一层池化的设计方式, 这一点和vgg16的前半部分很相似, 因此, 我在实现的过程中, 采用了vgg16的前10层。
网络设计
def vgg10_unet(self, input_shape=(256,256,3), weights='imagenet'):
vgg16_model = VGG16(input_shape=input_shape, weights=weights, include_top=False)
block4_pool = vgg16_model.get_layer('block4_pool').output
block5_conv1 = Conv2D(1024, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block4_pool)
block5_conv2 = Conv2D(1024, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block5_conv1)
block5_drop = Dropout(0.5)(block5_conv2)
block6_up = Conv2D(512, 2, activation='relu', padding='same', kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(block5_drop))
block6_merge = Concatenate(axis=3)([vgg16_model.get_layer('block4_conv3').output, block6_up])
block6_conv1 = Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block6_merge)
block6_conv2 = Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block6_conv1)
block6_conv3 = Conv2D(512, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block6_conv2)
block7_up = Conv2D(256, 2, activation='relu', padding='same', kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(block6_conv3))
block7_merge = Concatenate(axis=3)([vgg16_model.get_layer('block3_conv3').output, block7_up])
block7_conv1 = Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block7_merge)
block7_conv2 = Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block7_conv1)
block7_conv3 = Conv2D(256, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block7_conv2)
block8_up = Conv2D(128, 2, activation='relu', padding='same', kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(block7_conv3))
block8_merge = Concatenate(axis=3)([vgg16_model.get_layer('block2_conv2').output, block8_up])
block8_conv1 = Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block8_merge)
block8_conv2 = Conv2D(128, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block8_conv1)
block9_up = Conv2D(64, 2, activation='relu', padding='same', kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(block8_conv2))
block9_merge = Concatenate(axis=3)([vgg16_model.get_layer('block1_conv2').output, block9_up])
block9_conv1 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block9_merge)
block9_conv2 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block9_conv1)
block10_conv1 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(block9_conv2)
block10_conv2 = Conv2D(2, 1, activation='sigmoid')(block10_conv1)
model = Model(inputs=vgg16_model.input, outputs=block10_conv2)
return model
这样设计的好处就是, 我们前半部分采用了vgg网络可以使我们在训练网络的时候, 前半部分的权重, 我们可以加载vgg的预训练模型的权重来进行初始化, 当然, 你也可以在训练的时候, 冻结这几层网络, 只训练后半部分。
测试结果
感谢
学习的过程中, 参考了项目: https://github.com/zhixuhao/unet
论文引用: Convolutional Networks for Biomedical Image Segmentation
如果你看了这个项目对你有帮助, 麻烦帮我点颗星星, 谢谢