转置卷积的理解

作者: 苟且偷生小屁屁 | 来源:发表于2019-01-16 19:38 被阅读0次

转置卷积(Transposed convolution)用在什么地方？

转置卷积在图像的语义分割领域应用很广，如果说pooling层用于特征降维，那么在多个polling层后，就需要用转置卷积来进行分辨率的恢复。
比方说在全卷积神经网络中，up-sampling采用双线性插值进行分辨率的提升，而这种提升是非学习的，采用解卷积来完成上采样的工作，就可以通过学习的方式得到更高的精度。

转置卷积（解卷积）为什么叫“转置”？

对于下图，卷积核 $C$ 可以表示为：

图片.png

$\left( {\begin{array}{*{20}c} {\begin{array}{*{20}c} {w_{0,0} } & {w_{0,1} } & {w_{0,2} } & {\begin{array}{*{20}c} 0 & {w_{1,0} } & {w_{1,1} } & {\begin{array}{*{20}c} {w_{1,2} } & 0 & {w_{2,0} } & {\begin{array}{*{20}c} {w_{2,1} } & {w_{2,2} } & {\begin{array}{*{20}c} 0 & 0 & 0 & {\begin{array}{*{20}c} 0 & 0 \\ \end{array}} \\ \end{array}} \\ \end{array}} \\ \end{array}} \\ \end{array}} \\ \end{array}} \\ {\begin{array}{*{20}c} {\begin{array}{*{20}c} {\begin{array}{*{20}c} {\begin{array}{*{20}c} 0 & {w_{0,0} } & {w_{0,1} } \\ \end{array}} & {w_{0,2} } & 0 & {w_{1,0} } \\ \end{array}} & {w_{1,1} } & {w_{1,2} } & 0 \\ \end{array}} & {w_{2,0} } & {w_{2,1} } & {\begin{array}{*{20}c} {\begin{array}{*{20}c} {w_{2,2} } & 0 \\ \end{array}} & 0 & 0 & 0 \\ \end{array}} \\ \end{array}} \\ {\begin{array}{*{20}c} 0 & 0 & {\begin{array}{*{20}c} {\begin{array}{*{20}c} 0 & 0 & {w_{0,0} } & {w_{0,1} } \\ \end{array}} & {w_{0,2} } & 0 & {w_{1,0} } \\ \end{array}} & {\begin{array}{*{20}c} {w_{1,1} } & {w_{1,2} } & {\begin{array}{*{20}c} {\begin{array}{*{20}c} 0 & {w_{2,0} } \\ \end{array}} & {w_{2,1} } & {w_{2,2} } & 0 \\ \end{array}} \\ \end{array}} \\ \end{array}} \\ {\begin{array}{*{20}c} 0 & 0 & {\begin{array}{*{20}c} {\begin{array}{*{20}c} {\begin{array}{*{20}c} 0 & 0 & 0 \\ \end{array}} & {w_{0,0} } & {w_{0,1} } & {w_{0,2} } \\ \end{array}} & 0 & {w_{1,0} } & {\begin{array}{*{20}c} {\begin{array}{*{20}c} {w_{1,1} } & {w_{1,2} } \\ \end{array}} & 0 & {w_{2,0} } & {w_{2,1} } \\ \end{array}} \\ \end{array}} & {w_{2,2} } \\ \end{array}} \\ \end{array}} \right)$