Depthwise Separable Convolution

An efficient convolution structure that reduces CNN computation by separating spatial and channel transformations.

Depthwise separable convolution reduces computational burden significantly by splitting standard convolution into two stages. First, a spatial filter is applied separately to each channel; then cross-channel information is combined through pointwise convolution. It is a core building block of efficient CNN designs for mobile and embedded systems. It offers a strong balance between parameter efficiency and performance.