基于批归一化统计量的无源多领域自适应方法

Multi-source-free domain adaptation with batch normalization statistics

为解决传统的领域自适应方法训练期间源域数据并不总是可用这一问题,提出一种无源多领域自适应方法,有效完成当存在领域漂移现象时的图像分类任务。通过最小化源域和目标域数据的批归一化统计量距离减小域之间的分布差异,解决因无法访问源域数据而无法显式对齐源域与目标域的问题;采用基于近邻聚合策略的伪标签分类器辅助生成更加准确的伪标签,提高模型预测的准确性;通过学习最优的融合权重,将多个自适应后的源域模型进行有效融合。构建基于批归一化统计量的无源多领域自适应模型。性能对比试验和消融试验结果表明,与多个基线模型相比,本研究方法预测准确性提高0.6%~3.7%。

A multisourcefree domain adaptation method was proposed to solve the problem that the source data were not always available during the training of traditional domain adaptation methods,which effectively accomplished the image classification task with domain drift phenomenon.To solve this problem,a multisourcefree domain adaptation method was proposed to efficiently accomplish the image classification tasks in the presence of domain shift phenomenon.By minimizing the distance between the batch normalized statistics of the source domain and target domain data,the distribution difference between the fields was reduced.The problem that the source domain and target domain cannot be explicitly aligned due to the lack of access to the source domain data was solved.A pseudolabel classifier based on the nearest neighbor aggregation strategy was used to generate more accurate pseudolabels to improve the accuracy of model prediction.Multiple source domain models were efficiently combined after adaptation by learning the optimal fusion weights.A passive multidomain adaptive model based on batch normalized statistics was constructed.The results of performance comparison test and ablation test showed that compared with multiple baseline models,the prediction accuracy of this method was improved by 0.6%3.7%.