基于组合域的图像标签噪声预处理方法

Image label noise preprocessing method based on combined domain

数据集独立同分布(Independent and identically distributed,IID)的假设和数据集是否干净,分别对应组合域标签噪声领域数据集非独立同分布(Non-independent and identically distributed,Non-IID)和标签噪声问题。该文提出多标签图像多分类模型(Multi-label image multi-classification model,MIMM)预处理方法应对上述问题。为了缓和数据集独立同分布假设,该文证明组合域方法的迁移误差上界,并使用最大关键特征(Maximum key feature,MKF)函数和关键秩匹配分解(Key rank matching decomposition,KRMD)算法实现源域向组合同构数据的目标域迁移。为了保证数据集干净,引入干净的第三方数据微调模型参数,排除标签真实数量不确定导致训练集类别不平衡的情况。在Pascal/COCO数据集上进行实验,结果证明MIMM方法已达到准确率与运行效率的最佳平衡。

The assumption of independent and identically distributed(IID)datasets and the cleanliness of datasets respectively correspond to the non-independent and indentically distribution(Non-IID)and label noise in the combined domain label noise field.This paper proposes a preprocessing method based on the multi-label image multi-classification model(MIMM)to solve the above problems.In order to alleviate the assumption of independent and identically distributed datasets,this paper proves the upper bound of the transfer error of the combined domain method,and uses the maximum key feature(MKF)function and key rank matching decomposition(KRMD)algorithm to transfer the source domain to the target domain of combined isomorphic data.Finally,experimental results on Pascal/COCO datasets demonstrate that the MIMM method achieves the best balance between accuracy and efficiency.