Approach

• Depthwise Separable Convolution

  
  
  

  For MobileNets the depthwise convolution applies a single filter to each input channel. The pointwise convolution then applies a 1×1 convolution to combine the outputs the depthwise convolution.

Standard convolutions have the computational cost of: D K · D K · M · N · D F · D F
Depthwise separable convolutions cost: D K · D K · M · D F · D F + M · N · D F · D F
By expressing convolution as a two step process of filtering and combining we get a reduction in computation of:

• Network Structure

Experiment

References:
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications, Andrew G. Howard, 2017, arXiv