Liu Yongmin, Xiao Fengjiao, Zheng Xinying, Deng Weihao, Ma Haizhi, Su Xinyao, Wu Lei
School of Electronic Information and Physics, Central South University of Forestry Science and Technology, Changsha, 410004, China.
Research Center of Smart Forestry Cloudy, Central South Forestry University of Science and Technology, Changsha, 410004, China.
Sci Rep. 2025 Jan 8;15(1):1306. doi: 10.1038/s41598-025-85467-4.
Graph neural networks have excellent performance and powerful representation capabilities, and have been widely used to handle Few-shot image classification problems. The feature extraction module of graph neural networks has always been designed as a fixed convolutional neural network (CNN), but due to the intrinsic properties of convolution operations, its receiving domain is limited. This method has limitations in capturing global feature information and easily ignores key feature information of the image. In order to extract comprehensive and critical feature information, a new CA-MFE algorithm is proposed. The algorithm first utilizes different convolution kernels in CNN to extract multi-scale local feature information, and then based on the global feature extraction ability of attention mechanism, parallel processing of channel and spatial attention mechanism is used to extract multidimensional global feature information. This paper provides a comprehensive performance evaluation of the new model on both mini-ImageNet and tiered ImageNet datasets. Compared with the benchmark model, the classification accuracy has increased by 1.07% and 1.33% respectively; In the 5-way 5-shot task, the classification accuracy of the mini-ImageNet dataset was improved by 11.41%, 7.42%, and 5.38% compared to GNN, TPN, and dynamic models, respectively. The experimental results show that compared with the benchmark model and several representative Few-shot classification algorithm models, the new CA-MFE model has significant superior performance in processing few-shot classification data.
图神经网络具有优异的性能和强大的表征能力，已被广泛用于处理少样本图像分类问题。图神经网络的特征提取模块一直被设计为固定的卷积神经网络（CNN），但由于卷积操作的固有特性，其感受野有限。这种方法在捕获全局特征信息方面存在局限性，并且容易忽略图像的关键特征信息。为了提取全面且关键的特征信息，提出了一种新的CA-MFE算法。该算法首先利用CNN中的不同卷积核提取多尺度局部特征信息，然后基于注意力机制的全局特征提取能力，采用通道和空间注意力机制并行处理来提取多维全局特征信息。本文在mini-ImageNet和tiered ImageNet数据集上对新模型进行了全面的性能评估。与基准模型相比，分类准确率分别提高了1.07%和1.33%；在5-way 5-shot任务中，mini-ImageNet数据集的分类准确率相比GNN、TPN和动态模型分别提高了11.41%、7.42%和5.38%。实验结果表明，与基准模型和几种具有代表性的少样本分类算法模型相比，新的CA-MFE模型在处理少样本分类数据方面具有显著的优越性能。
