Boosting FasterNet: An Efficient PConv Block with Feature Reuse

The design of lightweight convolutional neural networks is crucial for deploying deep learning models on resourceconstrained devices. Early works such as MobileNet and ShuffleNet reduced computational costs by decomposing standard convolutions and restructuring network designs, yet they often introduced significant memory access overhead and fragmented operations, limiting their practical inference speed. Partial Convolution leverages feature sparsity to reduce computational complexity. However, it does not exploit the potential of sparse features and suffers from computational redundancy. Building upon Partial Convolution, this paper thoroughly investigates how to utilize feature sparsity more efficiently and proposes the Efficient PConv Block. This module maps extracted features to higher dimensions through linear layers and introduces a feature reuse mechanism to achieve dimensional expansion at minimal cost. Theoretical analysis and experiments demonstrate that the Efficient PConv Block significantly reduces parameter count and computational cost while achieving higher classification accuracy. Experimental results confirm that the Efficient PConv Block achieves a better balance between accuracy and efficiency. Furthermore, we incorporate BN and ReLU folding techniques to further optimize the inference speed of the Efficient PConv Block, making the network more suitable for deployment.