Deconvolution (Deconvolution)
Objective: To picture sampling, to obtain larger picture
mode: to fill empty picture
Source: FCN
- original conv without padding:
- original conv with padding:
Empty / expansion convolution (Dilated / Atrous Convolution)
Objective: To increase the receptive field without reducing the image size, instead of pooling and up-sampling
method: fill the cavity of convolution kernels receptive field exponential growth
Source: Dilated Residual Networks
Packet convolution (Group convolutions)
Objective: to reduce the amount of parameters
manner: each convolution of the input and convolved into k groups, respectively, k obtained in ouput group then splicing
Source: AlexNet
Depth separable convolution (Depthwise separable convolutions)
Objective: to reduce the amount of parameters
manner: correlation and channel spatial correlation separation
Source: Xception, MobileNet