1、已经存在的特征金字塔方法

为了检测到变化尺寸的目标,基于特征金字塔的检测器,在不同特征层之间,基于在k特征图上的决策,例如下图(a)所示,baseline检测器使用在特征层ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积上的特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_02

           ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_03

                                                             ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_04

其中ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_05。其中ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_06是骨干网络产生的特征图,ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_07是从后来的卷积层由底向上得到。ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_08代表了第l个卷积层执行的操作。ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_09代表检测子网络,通常采用一个单一的ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_10的卷积层来产生分类和框回归的输出。由于从金字塔层输入的深度不同,较浅的底层特征缺乏语义信息。

 

为了减少不同金字塔层之间的语义差距,有一些工作提出了使用横向连接的自顶向下结构,如图(c)所示。该结构通过横向连接将高层语义从顶层传播到底层,提高了分辨率,同时保持了空间上的高分辨率。第l层的特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_11产生的方式为

                                                                ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_12

其中ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_05ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_14是第l层的横向连接,ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_15是第l的自顶向下的连接。操作符ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_16代表两个特征图的组合,例如通道连接和相加。不同的方法仅仅采用了不同的ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_15ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_14

对特征金字塔这些方法比较抽象,他们依然有一些限制。因为自顶向下的连接以没有方向的方式传播语义,这些语义在各层上是不均匀分布的。结果是金字塔特征层之间的语义分隔依然存在。其次,这种特征的单向处理能力有限,无法生成丰富的上下文信息,从而提高所有尺度上的语义水平。为了解决这个问题我们开发了一个使用biLSTM在所有特征层之间以单向横向连接产生深度融合的语义。接下来的章节将展示我们提出方法的细节。

3.2、ScarfNet:整个结构

ScarfNet用两步来解决语义信息的不符:(1)、使用biLSTM来组合打散的语义信息。(2)、使用逐通道注意模块将融合的特征重新分布到每个特征层。整个结构如下图所示:

                                     ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_19

将第k个金字塔特征ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_07作为输入,ScarfNet产生新的第l个特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_11为:

                                       ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_22

其中ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_05,如式(6)所示ScarfNet由两部分组成:语义重组网络(ScNet)和注意重分布网络(ArNet):

  1. ScNet通过biLSTM来融合金字塔特征ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_07,并且用融合的语义产生输出特征。
  2. ArNet收集从biLSTM的输出特征,并且用逐通道注意力来产生高质量的语义多尺度特征,连接到原始的特征金子塔上。最终,结果特征图用检测子网络ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_09单独处理来产生最终的检测结果。

3、语义组合网络(ScNet)

通过ScNet产生的特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_26为:

                                     ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_27

ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_28是第l层的输出特征图,细节如下图所示,描述了ScNet的细节。

ScNet使用biLSTM在不同的金字塔之间均匀的融合打散的特征。

biLSTM通过门函数,在多尺度层上选择融合语义信息。

ScNet由匹配模块和biLSTM组成。

匹配模块首先对金字塔特征ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_07的尺寸进行变换,使他们的尺寸相同。

然后使用ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_30的卷积层来调整通道维度。

结果,匹配模块产生通道数和尺寸都相同的特征图。尺寸变换操作通过双线性插值来完成。

biLSTM和参考文献[23]相同。基于全局池化的结果,对输入连接和门参数的计算使用卷积层,来显著的节省计算。

                                      ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_31

特别地,biLSTM的操作可以简化为:

                                   ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_32

其中ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_操作符_33代表哈达玛积,biLSTM的状态在前向和后向都更新。上式为前向更新,后向更新的表达式类似。

4、注意力重分布网络(ArNet)

ArNet产生高层的语义特征图,连接到原始的金字塔特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_02上,表达式为:

                              ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_35

操作符ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_16代表逐通道连接。ArNet的具体结构下图所示。ArNet连接biLSTM的输出ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_卷积_26,对他们应用逐通道注意力机制。注意力机制的权重通过构建ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_30的向量获得,具体方式为使用全局平均池化,并且将将它传递到两个全连接层,最后再接一个sigmoid函数。注意,这些逐通道注意力模块允许选择将语义传播到金字塔的每层。一旦注意力的权重使用了,匹配模块将特征图的结果进行下采样,并且应用ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_自顶向下_30的卷积来匹配通道维数,利用这些原始的金字塔特征。最终,输出的匹配模块连接到原始的特征图ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_权重_02上,来产生高的语义特征ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_11

                              ScarfNet: Multi-scale Features with Deeply Fused and Redistributed Semantics_语义信息_42