MobileNets: Efficient Convolutional Neural Networks for Mobile Vision .

Abstract

MobileNet uses depthwise separable convolutions to build lightweight models for mobile and embedded devices.

Key Innovation

Depthwise separable convolution:

• Depthwise: Apply single filter per input channel
• Pointwise: 1x1 convolution to combine

This reduces computation by 8-9x compared to standard convolutions.

Impact

Enabled real-time computer vision on smartphones and edge devices. Widely used in mobile apps for object detection, segmentation, etc.