基于困难样本对激励的小样本图像分类方法

A Few-Shot Image Classification Method by Hard Pairwise-Based Excitation

使用少量标签样本训练得到的传统模型往往预测精度低、泛化能力弱,很难应用到实际生产中.针对小样本图像提出一种基于困难样本对激励分类方法,包括预训练阶段和元学习阶段.预训练阶段在基类数据集上训练编码器,并作为元学习阶段的初始特征编码器;元学习阶段将进一步优化此编码器,元训练过程使用本质特征法降低异常样本对质心的影响;结合度量学习与元学习设计了困难样本对激励损失函数,从样本对角度出发,在训练过程中引导模型扩大正负样本间距离,使同类样本更加紧凑.在公开数据集mini-ImageNet,tiered-ImageNet上进行实验的结果表明,分类精度分别为64.12%,70.15%,验证了所提方法的有效性和可行性.

Traditional models trained with a small number of label samples often have low prediction accuracy and weak generalization ability and are difficult to be applied to practical production.A classification method named hard pairwise-based excitation is proposed for the few-shot image classification,including pre-training stage and meta learning stage.Pre-training stage trains the encoder on the base class dataset and used as the initial feature encoder in the meta learning stage;In the meta learning stage,the encoder will be further optimized,and the meta training process uses the essential feature method to reduce the impact of abnormal samples on the centroid;Combining measurement learning and meta learning,a loss function named hard sample-pairs excitation is designed.From the perspective of sample pairs,the model is guided to expand the distance between positive and negative samples during the training process,making similar samples more compact.The experimental results on public datasets mini-ImageNet and tiered-ImageNet show that the classification accuracy is 64.12%and 70.15%,respectively,verifying the effectiveness and feasibility of the proposed method.