voc-fcn-alexnet网络结构理解

一、写在前面

fcn是首次使用cnn来实现语义分割的，论文地址：fully convolutional networks for semantic segmentation

实现代码地址：https://github.com/shelhamer/fcn.berkeleyvision.org

全卷积神经网络主要使用了三种技术：

1. 卷积化（Convolutional）

2. 上采样（Upsample）

3. 跳跃结构（Skip Layer）

为了便于理解，我拿最简单的结构voc-fcn-alexnet进行说明，该网络结构主要用到了前面两个技术，不包含跳跃结构。

二、voc-fcn-alexnet 的train.prototxt文件

layer {
  name: "data"
  type: "Python"
  top: "data"
  top: "label"
  python_param {
    module: "voc_layers"
    layer: "SBDDSegDataLayer"
    param_str: "{\'sbdd_dir\': \'../data/sbdd/dataset\', \'seed\': 1337, \'split\': \'train\', \'mean\': (104.00699, 116.66877, 122.67892)}"
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size:
    stride:
  }
}
layer {
  name: "norm1"
  type: "LRN"
  bottom: "pool1"
  top: "norm1"
  lrn_param {
    local_size:
    alpha: 0.0001
    beta: 0.75
  }
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "norm1"
  top: "conv2"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu2"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size:
    stride:
  }
}
layer {
  name: "norm2"
  type: "LRN"
  bottom: "pool2"
  top: "norm2"
  lrn_param {
    local_size:
    alpha: 0.0001
    beta: 0.75
  }
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "norm2"
  top: "conv3"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu3"
  type: "ReLU"
  bottom: "conv3"
  top: "conv3"
}
layer {
  name: "conv4"
  type: "Convolution"
  bottom: "conv3"
  top: "conv4"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu4"
  type: "ReLU"
  bottom: "conv4"
  top: "conv4"
}
layer {
  name: "conv5"
  type: "Convolution"
  bottom: "conv4"
  top: "conv5"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu5"
  type: "ReLU"
  bottom: "conv5"
  top: "conv5"
}
layer {
  name: "pool5"
  type: "Pooling"
  bottom: "conv5"
  top: "pool5"
  pooling_param {
    pool: MAX
    kernel_size:
    stride:
  }
}
layer {
  name: "fc6"
  type: "Convolution"
  bottom: "pool5"
  top: "fc6"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  bottom: "fc6"
  top: "fc6"
}
layer {
  name: "drop6"
  type: "Dropout"
  bottom: "fc6"
  top: "fc6"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "fc7"
  type: "Convolution"
  bottom: "fc6"
  top: "fc7"
  convolution_param {
    num_output:
    pad:
    kernel_size:
    group:
    stride:
  }
}
layer {
  name: "relu7"
  type: "ReLU"
  bottom: "fc7"
  top: "fc7"
}
layer {
  name: "drop7"
  type: "Dropout"
  bottom: "fc7"
  top: "fc7"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "score_fr"
  type: "Convolution"
  bottom: "fc7"
  top: "score_fr"
  param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  convolution_param {
    num_output:
    pad:
    kernel_size:
  }
}
layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult:
  }
  convolution_param {
    num_output:
    bias_term: false
    kernel_size:
    stride:
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore"
  bottom: "data"
  top: "score"
  crop_param {
    axis:
    offset:
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "score"
  bottom: "label"
  top: "loss"
  loss_param {
    ignore_label:
    normalize: true
  }
}

三、网络结构

假设输入的图片为500x500，

根据train.prototxt文件，可以得到上图的网络结构，该网络结构除了前五层的卷积层，也把后面的三层改为了卷积层，score_fr是卷积层的最后一层，也叫heatmap热图，热图就是我们最重要的高维特诊图，得到高维特征的heatmap之后，就是最重要的一步也是最后的一步，对原图像进行upsampling（即反卷积），把图像进行放大，得到原图像的大小。

四、损失函数

该网络的损失函数为SoftmaxWithLoss。首先进行softmax求解，求出每个像素点属于不同类别的概率，因为总共是分为21类，所以每个像素点对应21个概率值（输出通道数为21）。然后求解每个像素点所属实际类别概率的log值之和的平均，再取负数，可得到损失函数，参考如下：

end