LIU Yongmin, XIAO Fengjiao, QIAO mengyuan, DENG Weihao, MA Haizhi
Control Engineering of China. 2025, 32(6): 1008-1015.
Because the general graph neural network feature extraction module is designed as a fixed convolutional neural network (CNN), which leads to the existence of limited acceptance domain and easy to ignore the key feature information of the image when capturing the global feature information. In order to extract comprehensive and critical feature information, a novel CA-MFE algorithm is proposed: firstly, different convolutional kernels in the CNN are utilized to capture multi-scale local feature information, and then the channel and spatial attention mechanisms are processed in parallel according to the global feature extraction capability of the attention mechanism, in order to extract multi-dimensional global feature information. The performance of the new model is evaluated on mini-ImageNet and tiered-ImageNet datasets for a comprehensive and comprehensive performance evaluation. The classification accuracies are improved by 1.07% and 1.33% compared to the baseline model, respectively. Using the mini-ImageNet dataset in the 5-way 5-shot task, the classification accuracies are improved by 11.41%, 7.42%, and 5.38% compared to the GNN, TPN, and Dynamic models, respectively. The experimental results show that the CA-MFE model is significantly superior to the baseline model and several representative small-sample classification algorithms when dealing with small-sample classification data.