多类(4类)图像分割的UNet模型实现

Question

我试图在标记的图像数据上实现一个UNet模型。数据集包含大约10,000幅图像及其各自的掩码(有色-RGB)。

图像尺寸: 500×500×3

面具不是黑白的，它们是彩色的(RGB)，有3个类(技术上是4个)：

• 背景:黑色
• 第1类:红色
• 第2类:绿色
• 第三班:蓝色

这是模型的最后两个CONV块的代码：

    model = Conv2D(64,(3,3),strides=(1, 1),padding='same')(concat_5)
    model = LeakyReLU(0.1)(model)
    model = BatchNormalization()(model)

    model = Conv2D(4,(3,3),strides=(1, 1),padding='same', activation="softmax")(model)

    model = Model(base_model.input,model)

模型架构：

Model: "functional_1"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 500, 500, 3) 0                                            
__________________________________________________________________________________________________
block1_conv1 (Conv2D)           (None, 500, 500, 64) 1792        input_1[0][0]                    
__________________________________________________________________________________________________
block1_conv2 (Conv2D)           (None, 500, 500, 64) 36928       block1_conv1[0][0]               
__________________________________________________________________________________________________
block1_pool (MaxPooling2D)      (None, 250, 250, 64) 0           block1_conv2[0][0]               
__________________________________________________________________________________________________
block2_conv1 (Conv2D)           (None, 250, 250, 128 73856       block1_pool[0][0]                
__________________________________________________________________________________________________
block2_conv2 (Conv2D)           (None, 250, 250, 128 147584      block2_conv1[0][0]               
__________________________________________________________________________________________________
block2_pool (MaxPooling2D)      (None, 125, 125, 128 0           block2_conv2[0][0]               
__________________________________________________________________________________________________
block3_conv1 (Conv2D)           (None, 125, 125, 256 295168      block2_pool[0][0]                
__________________________________________________________________________________________________
block3_conv2 (Conv2D)           (None, 125, 125, 256 590080      block3_conv1[0][0]               
__________________________________________________________________________________________________
block3_conv3 (Conv2D)           (None, 125, 125, 256 590080      block3_conv2[0][0]               
__________________________________________________________________________________________________
block3_pool (MaxPooling2D)      (None, 62, 62, 256)  0           block3_conv3[0][0]               
__________________________________________________________________________________________________
block4_conv1 (Conv2D)           (None, 62, 62, 512)  1180160     block3_pool[0][0]                
__________________________________________________________________________________________________
block4_conv2 (Conv2D)           (None, 62, 62, 512)  2359808     block4_conv1[0][0]               
__________________________________________________________________________________________________
block4_conv3 (Conv2D)           (None, 62, 62, 512)  2359808     block4_conv2[0][0]               
__________________________________________________________________________________________________
block4_pool (MaxPooling2D)      (None, 31, 31, 512)  0           block4_conv3[0][0]               
__________________________________________________________________________________________________
block5_conv1 (Conv2D)           (None, 31, 31, 512)  2359808     block4_pool[0][0]                
__________________________________________________________________________________________________
block5_conv2 (Conv2D)           (None, 31, 31, 512)  2359808     block5_conv1[0][0]               
__________________________________________________________________________________________________
block5_conv3 (Conv2D)           (None, 31, 31, 512)  2359808     block5_conv2[0][0]               
__________________________________________________________________________________________________
block5_pool (MaxPooling2D)      (None, 15, 15, 512)  0           block5_conv3[0][0]               
__________________________________________________________________________________________________
conv2d_transpose_5 (Conv2DTrans (None, 31, 31, 256)  1179904     block5_pool[0][0]                
__________________________________________________________________________________________________
leaky_re_lu_10 (LeakyReLU)      (None, 31, 31, 256)  0           conv2d_transpose_5[0][0]         
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 31, 31, 256)  1024        leaky_re_lu_10[0][0]             
__________________________________________________________________________________________________
concatenate_5 (Concatenate)     (None, 31, 31, 768)  0           batch_normalization_10[0][0]     
                                                                 block5_conv3[0][0]               
__________________________________________________________________________________________________
conv2d_6 (Conv2D)               (None, 31, 31, 512)  3539456     concatenate_5[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_11 (LeakyReLU)      (None, 31, 31, 512)  0           conv2d_6[0][0]                   
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 31, 31, 512)  2048        leaky_re_lu_11[0][0]             
__________________________________________________________________________________________________
conv2d_transpose_6 (Conv2DTrans (None, 62, 62, 512)  2359808     batch_normalization_11[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_12 (LeakyReLU)      (None, 62, 62, 512)  0           conv2d_transpose_6[0][0]         
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 62, 62, 512)  2048        leaky_re_lu_12[0][0]             
__________________________________________________________________________________________________
concatenate_6 (Concatenate)     (None, 62, 62, 1024) 0           batch_normalization_12[0][0]     
                                                                 block4_conv3[0][0]               
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 62, 62, 512)  4719104     concatenate_6[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_13 (LeakyReLU)      (None, 62, 62, 512)  0           conv2d_7[0][0]                   
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 62, 62, 512)  2048        leaky_re_lu_13[0][0]             
__________________________________________________________________________________________________
conv2d_transpose_7 (Conv2DTrans (None, 125, 125, 512 2359808     batch_normalization_13[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_14 (LeakyReLU)      (None, 125, 125, 512 0           conv2d_transpose_7[0][0]         
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 125, 125, 512 2048        leaky_re_lu_14[0][0]             
__________________________________________________________________________________________________
concatenate_7 (Concatenate)     (None, 125, 125, 768 0           batch_normalization_14[0][0]     
                                                                 block3_conv3[0][0]               
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 125, 125, 256 1769728     concatenate_7[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_15 (LeakyReLU)      (None, 125, 125, 256 0           conv2d_8[0][0]                   
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 125, 125, 256 1024        leaky_re_lu_15[0][0]             
__________________________________________________________________________________________________
conv2d_transpose_8 (Conv2DTrans (None, 250, 250, 256 590080      batch_normalization_15[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_16 (LeakyReLU)      (None, 250, 250, 256 0           conv2d_transpose_8[0][0]         
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 250, 250, 256 1024        leaky_re_lu_16[0][0]             
__________________________________________________________________________________________________
concatenate_8 (Concatenate)     (None, 250, 250, 384 0           batch_normalization_16[0][0]     
                                                                 block2_conv2[0][0]               
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 250, 250, 128 442496      concatenate_8[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_17 (LeakyReLU)      (None, 250, 250, 128 0           conv2d_9[0][0]                   
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 250, 250, 128 512         leaky_re_lu_17[0][0]             
__________________________________________________________________________________________________
conv2d_transpose_9 (Conv2DTrans (None, 500, 500, 128 147584      batch_normalization_17[0][0]     
__________________________________________________________________________________________________
leaky_re_lu_18 (LeakyReLU)      (None, 500, 500, 128 0           conv2d_transpose_9[0][0]         
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 500, 500, 128 512         leaky_re_lu_18[0][0]             
__________________________________________________________________________________________________
concatenate_9 (Concatenate)     (None, 500, 500, 192 0           batch_normalization_18[0][0]     
                                                                 block1_conv2[0][0]               
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 500, 500, 64) 110656      concatenate_9[0][0]              
__________________________________________________________________________________________________
leaky_re_lu_19 (LeakyReLU)      (None, 500, 500, 64) 0           conv2d_10[0][0]                  
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 500, 500, 64) 256         leaky_re_lu_19[0][0]             
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 500, 500, 4)  2308        batch_normalization_19[0][0]     
==================================================================================================
Total params: 31,948,164
Trainable params: 31,941,892
Non-trainable params: 6,272
__________________________________________________________________________________________________

当我训练模型时，我会得到以下错误：

InvalidArgumentError:  Incompatible shapes: [4,4,500,500] vs. [4,3,500,500]
     [[node categorical_crossentropy/mul (defined at <ipython-input-15-b60b11ed9e76>:9) ]] [Op:__inference_train_function_5874]

Function call stack:
train_function

模型的最后一层的维数似乎存在一些问题。我想就以下三件事提出建议：

1. 最后一层的维数(考虑4类的多类分割)
2. 我应该在最后一层使用哪种激活函数？目前，我正在使用softmax激活。
3. 在编译模型时，我应该使用什么样的损失？目前，我正在使用categorical_crossentropy。

Answer 1

我认为问题来自你的课堂。它们的格式是什么？它们是单热编码的吗？

你的激活和丢失功能对我来说是正确的。我认为上一个卷积层应该使用1x1大小的过滤器(类似于最初的纸)。它确保了更好的像素分类。