Pyramid Feature Attention Network for Saliency detection 论文解读

Pyramid Feature Attention Network for Saliency detection //CVPR 2019论文解读

提出的问题:

Pooling层在降低feature map的同时,显著物体的边界也变得模糊

现有解决方法:

  1. 引入手工特征来保留边界
  2. 整合multi-level+multi-scale特征,深层包含全局上下文特征(准确定位显著物体位置),浅层包含空间结构特征(定位边缘)

存在的问题:

  1. 难以融合这些独立提取的特征,同时这是一个费时的课程
  2. 不加差别的融合这些high-level和low-level特征,导致次优效果;另外,gate function和progressive attention通常从单纯的一个方向来选择特征,并且忽视了高低级特征的差别

Contribution:

  1. 提出Pyramid Feature Attention (PFA) Net. 考虑到low-level特征图包含许多noises,采用spatial attention module模型过滤背景细节;high-level特征图仅仅包含了一个大致的显著性区域,采用context-aware pyramid feature extraction module + CA module 获取上下文信息
  2. 审计了一个新的edge preservation loss,引导网络在边界定位中学习更详细的信息
  3. 实验效果- - -

Pyramid Feature Attention Network for Saliency detection 论文解读流程:
用CPFE获取多尺度和多感受野获取high-level特征,用CA对不同channel分配不同权重(Figure1. f);
(Figure1. c)low-level特征图包含许多noises,不是所有边缘信息都有助于细化显著结果,选择更加关注于显著物体与背景之间的边界,用SA更好的关注有效的low-levle特征,获得更清楚的边界;最后再融合低高级特征。

框架

Pyramid Feature Attention Network for Saliency detection 论文解读

CPFE

Pyramid Feature Attention Network for Saliency detection 论文解读

CA+SA

Pyramid Feature Attention Network for Saliency detection 论文解读Pyramid Feature Attention Network for Saliency detection 论文解读
Pyramid Feature Attention Network for Saliency detection 论文解读
Pyramid Feature Attention Network for Saliency detection 论文解读
Pyramid Feature Attention Network for Saliency detection 论文解读

Edge preservation loss

交叉墒:
Pyramid Feature Attention Network for Saliency detection 论文解读
Laplace 算子:
Pyramid Feature Attention Network for Saliency detection 论文解读
Pyramid Feature Attention Network for Saliency detection 论文解读最终loss:
Pyramid Feature Attention Network for Saliency detection 论文解读

实验结果

Pyramid Feature Attention Network for Saliency detection 论文解读
Pyramid Feature Attention Network for Saliency detection 论文解读Pyramid Feature Attention Network for Saliency detection 论文解读Pyramid Feature Attention Network for Saliency detection 论文解读