高级检索

    基于多尺度标签传播的小样本图像分类

    Few-Shot Image Classification Based on Multi-Scale Label Propagation

    • 摘要: 在小样本条件下,由于低数据问题,即标记数据较少且难以收集,采用传统的深度学习很难训练出一个好的分类器.最近的研究中,基于低维局部信息度量方法和标签传播网络(transductive propagation network, TPN)算法取得了较好的分类效果,并且局部信息可以很好地度量特征与特征之间的关系,但是低数据问题依然存在.为了解决低数据问题,提出基于多尺度的标签传播网络(multi-scale label propagation network, MSLPN)方法,其核心思想在于利用多尺度生成器生成多个尺度的图像特征,通过关系度量模块获得多个不同尺度特征下的样本相似性得分,并通过集成不同尺度的相似性得分获得分类结果,具体地,方法首先通过多尺度生成器生成不同尺度的图像特征,然后利用多尺度信息的相似性得分进行标签传播,最后通过多尺度标签传播结果计算获得分类结果.与TPN相比,在数据集miniImageNet上,5-way 1-shot和5-way 5-shot设置中的分类准确率分别提高了2.77%和4.02%;在数据集tieredImageNet上,5-way 1-shot和5-way 5-shot设置中分类准确率分别提高了1.16%和1.27%.实验结果表明,利用多尺度特征信息可有效提高分类准确率.

       

      Abstract: Under the condition of few-shot, due to the problem of low data, in other words, the labeled data is rare and difficult to gather, it is very difficult to train a good classifier by traditional deep learning. In recent researches, the method based on measuring low level local information and TPN(transductive propagation network) has achieved good classification results. Moreover, local information can measure the relation between features well, but the problem of low data still exists. In order to solve the issue of low data, MSLPN (multi-scale label propagation network) based on TPN is proposed in this paper. The core idea of the method is to use a multi-scale generator to generate image features of multiple scales, obtain the similarity scores of samples with different scale features through the relational measurement module, and obtain classification results by integrating similarity scores at different scales. Specifically, the method firstly generates multiple image features of different scales through a multi-scale generator. And then, the similarity scores of the multi-scale information are used for label propagation. Finally, classification results are obtained by calculating the multi-scale label propagation results. Compared with TPN, in miniImageNet, the classification accuracy of 5-way 1-shot and 5-way 5-shot settings is increased by 2.77% and 4.02% respectively. While in tieredImageNet, the classification accuracy of 5-way 1-shot and 5-way 5-shot settings is increased by 1.16% and 1.27% respectively. The experimental results show that the proposed method in this paper can effectively improve the classification accuracy by using multi-scale feature information.

       

    /

    返回文章
    返回