基于样本选择的标签含噪图像分类  

Label Noisy Image Classification Based on Sample Selection

作　　者：闻铮 曹国[1] WEN Zheng;CAO Guo(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机系统应用》2024年第2期54-61,共8页Computer Systems & Applications

基　　金：国家自然科学基金(62201282);江苏省自然科学基金(BK20231456)。

摘　　要：标签噪声广泛存在、无法避免且影响深度网络模型的性能.利用神经网络的“记忆效应”,基于小损失原则的样本选择方法能简单有效地处理标签噪声.本文基于特征空间中样本距离越近越相似的原则,结合样本的高低置信度假设,提出了新的样本选择原则以及二阶段加权样本选择重标签方法(WSSR-2s).(1)在训练前期阶段,对于高置信度样本,在特征空间中对其票权进行加权,更好地引导训练;(2)在训练中后期阶段,对于低置信度样本,将其票权转移给其最相似的特征样本,以更正确地训练.在合成噪声数据集CIFAR-10、CIFAR-100以及真实噪声数据集ANIMAL-10N、WebVision的实验结果表明,本文提出的方法取得更高的精度,能够更好地处理标签噪声问题.Label noise is widely present and unavoidable,and it affects the performance of deep network models.Sample selection methods based on the principle of small loss can easily and effectively handle label noise by the“memory effect”of neural networks.This study proposes a new sample selection principle and a two-stage weighted sample selection and relabeling method(WSSR-2s)based on the principle that a closer sample distance in the feature space results in more similarity,combined with the assumption of high and low confidence of the samples.In the early training stage,for high-confidence samples,their voting rights are weighted in the feature space to better guide training.In the middle and later stages of training,for low-confidence samples,their voting rights are transferred to their most similar feature samples for more accurate training.The experimental results on synthetic noise datasets CIFAR-10 and CIFAR-100,as well as real noise datasets ANIMAL-10N and WebVision,show that the proposed method achieves higher accuracy and can better handle label noise problems.

关 键 词：标签噪声 样本选择 置信度假说 样本票权 样本加权 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]