《Visual Saliency Detection Based on Full Convolution Neural Networks and Center Prior》阅读理解

作者：Muwei Jian， Jiaojin Wang ， Xiangyu Liu，Hui Yu

论文链接：Visual Saliency Detection Based on Full Convolution Neural Networks and Center Prior | IEEE Conference Publication | IEEE Xplore

2019 12th International Conference on Human System Interaction (HSI)

论文提出了基于FCN 和 center prior 的 VSOD 方法。目的是解决 the challenge in exploiting the consistency of spatiotemporal information。论文方法包括三个步骤：First，用FCN提取显著性特征。Second，通过weight sharing strategy扩展注意力特征。Finally，通过中心先验center prior来优化显著性结果。some example 如图1所示。

 网络结构如下：

 Deep networks for frame saliency

 通过FCNs完成。

Weighting share strategy

权重共享如下图所示：

 the vector in the deep network is thus given by:

 其在网络中的应用表示为：

 Final Saliency refinement

 采用center prior 进行细化。结果如图1所示。

 EXPERIMENTAL RESULTS