lartpang / Machine-Deep-Learning

:wave: ML/DL学习笔记(基础+论文)
262 stars 58 forks source link

论文-R3Net: Recurrent Residual Refinement Network for Saliency Detection #38

Closed ghost closed 5 years ago

ghost commented 5 years ago

image

image

ghost commented 5 years ago
  1. 使用ResNeXt提取特征, 分成了五个阶段 image image
  2. 使用前三个阶段产生的特征图上采样到输入图片(边长)的四分之一(也就是第一阶段卷积输出的特征图大小)后拼接在一起, 卷积融合并减少特征数,产生低级融合特征 image
    # followed by 3 PReLU activation functions[He et al., 2015]
    self.reduce_low = nn.Sequential(
    nn.Conv2d(64 + 256 + 512, 256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.PReLU(),
    nn.Conv2d(256, 256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.PReLU(),
    nn.Conv2d(256, 256, kernel_size=1), nn.BatchNorm2d(256), nn.PReLU()
    )

    PRuLU

  3. 使用最后的第4/5阶段的输出特征图上采样(放缩到第4阶段输出的特征图大小)拼接后卷积获得高级融合特征 image
    self.reduce_high = nn.Sequential(
    nn.Conv2d(1024 + 2048, 256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.PReLU(),
    nn.Conv2d(256, 256, kernel_size=3, padding=1), nn.BatchNorm2d(256), nn.PReLU(),
    _ASPP(256)
    )
  4. 从高级融合特征中预测初始显著性图(S0), 然后开始进入一系列的RRBs结构中进行迭代 image
  5. 在最终的输出作为最终的显著性图预测
lartpang commented 5 years ago

image

lartpang commented 5 years ago

image