CNN（画像認識・UNet）の構造に関する質問

kerasを用いて、Unetを作成しています。
構造は下記のような感じだと思いますが、参考にしているコードがUnetじゃないのではないかと疑っています。
自分の認識だと、各層のユニット数が、入力層から徐々に少なくなって、中間層の真ん中あたりで最小になって、出力層に向かって増えていくイメージでした。
しかし、参考にしているコードは、中間層の真ん中あたりで、ユニット数が最大になっています。なぜでしょうか？

Python
1def unet(pretrained_weights = None,input_size = (256,256,1)):
2    inputs = Input(input_size)
3    conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
4    conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv1)
5    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
6    conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
7    conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv2)
8    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
9    conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
10    conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv3)
11    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
12    conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
13    conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv4)
14    drop4 = Dropout(0.5)(conv4)
15    pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)
16
17    conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
18    conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv5)
19    drop5 = Dropout(0.5)(conv5)
20
21    up6 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(drop5))
22    merge6 = concatenate([drop4,up6], axis = 3)
23    conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge6)
24    conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv6)
25
26    up7 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv6))
27    merge7 = concatenate([conv3,up7], axis = 3)
28    conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge7)
29    conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv7)
30
31    up8 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv7))
32    merge8 = concatenate([conv2,up8], axis = 3)
33    conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge8)
34    conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv8)
35
36    up9 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv8))
37    merge9 = concatenate([conv1,up9], axis = 3)
38    conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge9)
39    conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)
40    conv9 = Conv2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)
41    conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)
42
43    model = Model(input = inputs, output = conv10)
44
45    model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', metrics = ['accuracy'])
46    
47    #model.summary()
48
49    if(pretrained_weights):
50        model.load_weights(pretrained_weights)
51
52    return model