A Lightweight Convolutional Network Based on RepLKnet and Ghost Convolution
Feiyang Xue, Lan Pang, Xiying Huang, Ruixiong Ma · 2024
Over the past two years, convolutional neural networks have witnessed a resurgence in the domain of computer vision, owing to the contemporary design of large convolutional kernels and the corresponding optimization techniques. Nevertheless, these large kernel convolutional neural networks primarily concentrate on enlarging the size of the convolutional kernel, while overlooking the significant amount of redundant information that emerges among the feature map channels as the convolutional kernel expands. With the aim of enhancing the computational efficiency of large convolutional kernel networks and enabling their application in more resource-constrained settings, this paper incorporates ghost convolution to modify the RepLK block, thereby effectively exploiting these redundancies. The validation outcomes on the CIFAR-100 dataset indicate that the lightweight and enhanced network model attains 4.32% of the FLOPs and 4.77% of the parameter counts of the RepLKnet, accompanied by merely a 4.04% reduction in Top1 accuracy and a 0.15% decrease in Top5 accuracy.