GhostNet: More Features From Cheap Operations

GhostNet addresses the challenge of deploying convolutional neural networks on embedded devices, where memory and computation resources are limited. The authors note that redundancy in feature maps is an important characteristic of successful CNNs, yet it has rarely been investigated in architecture design. They propose a Ghost module that generates more feature maps from cheap operations: beginning with a set of intrinsic feature maps, it applies a series of low-cost linear transformations to produce many additional ghost feature maps that reveal the information underlying the intrinsic features.

The Ghost module works as a plug-and-play component to upgrade existing CNNs, and stacking Ghost modules into Ghost bottlenecks yields the lightweight GhostNet. On the ImageNet ILSVRC-2012 classification benchmark, GhostNet achieves higher recognition performance, for example 75.7% top-1 accuracy, than MobileNetV3 at similar computational cost, showing that the Ghost module is an effective alternative to standard convolution layers. The code is released publicly.