Enhanced Module design based on ShuffleNetV2
Sifeng Luo, Huang Huang, Xi Huang · 2023
Among the general lightweight deep neural network models, ShuffleNetV2 is a portable and efficient model. To further make the network lighter and more powerful in recognition, an Enhanced Shuffle Block (ESBlock) is proposed, which is a symmetric module consisting of a depth-wise separable convolutional superposition to replace ShuffleNetV2's fundamental module. Then, by re-optimizing the configuration of channels, the expansion factor within the basic block, the layer of each stage, it is possible to improve the performance. And a Portable Attention Module (PAM) is added to enhance feature extraction and model's utilization capability. Experimental findings demonstrate that the enhanced network accomplishes 94.15% and 73.52% recognition accuracy on Cifar10 and Cifar100 datasets with only 0.66 M and 0.77 M parameters, respectively. Compared with the ShuffleNetV2, the accuracy is increased by 1% and 0.75%, and parameters is decreased by 50% and 44% respectively.