A Few-shot Image Classification Model Based on Two-stage Feature Space Enhancement
LI Gexian
ZHANG Xiaoshuang
HE Yongjiao
DU Yang
ZHANG Yansha
WANG Lin
Abstract：In few-shot learning tasks,a few-shot image classification model based on two-stage feature enhancement was proposed to address the issue that traditional backbone convolutional networks,due to the loss of feature information caused by the neglect of detailed features in multi-layer convolutions,resulted in low image classification accuracy.Firstly,this model was introduced with a median-enhanced spatial and channel attention block(MESC)in the lower layers of the residual network(ResNet)12.Secondly,this model was introduced with a spatial group-wise enhance(SGE)module in the middle and upper layers of the ResNet12 to improve the ability of semantic feature learning in convolutional neural networks and enable the model to effectively extract key information from feature maps.The model enhanced the feature representation of limited training samples to improve classification performance and enhance the model's robustness to noise.The results showed that,on the California Institute of Technology-University of California at San Diego birds(CUB)-200-2011 dataset,the classification accuracy of this model were respectively improved by about 5.15%and 1.92%compared with the distribution propagation graph network(DPGN)model under the two parameter settings of 5-way 1-shot and 5-way 5-shot.On the tiered ImageNet(tieredImageNet)dataset,the classification accuracy was respectively increased by about 1.04%and 0.55%compared with the DPGN model under these two parameter settings.The performance of the few-shot image classification task was significantly improved by this model.
Keywords：convolutional neural networkfew-shot learningimage classificationfeature space enhancementattention mechanismchannel attentionspatial group enhancement
Publication Date：2025-06-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：7( 231-236,279 )