GhostPAModule: Optimizing Cheap Operation via Feature Enhancement

This paper proposes improvement methods for the module GhostModule in lightweight neural networks, which has limited feature representation capabilities. We propose an improved GhostPAModule module. The method is to add pointwise convolution and attention mechanism after the cheap operation branch of GhostModule, in order to model the dependencies between features and calibrate the features. We validate this module on the CIFAR-10 dataset, and the results show that GhostPAModule achieves a 0.94% improvement in classification accuracy compared to the original GhostModule. This demonstrates the importance of modeling feature dependencies and feature calibration for improving the representation capabilities of lightweight networks.