上采样通常有四种方式:
-
upsample / Interpolation
-
unpooling
-
deconvolution / transpose convolution
-
pixel shuffle / sub-pixel convolution
-
反卷积/转置卷积:原理上是转置、逆变换,形式上是先padding再卷积,使得输出分辨率变大:https://github.com/vdumoulin/conv_arithmetic、https://blog.csdn.net/qq_16234613/article/details/79387345、https://blog.csdn.net/Fate_fjh/article/details/52882134
-
pixel_shuffle/sub-pixel conv:https://blog.csdn.net/g11d111/article/details/82855946、https://blog.csdn.net/shwan_ma/article/details/78440394
网友评论