基于上下文一致性改进的目标分类方法研究

Research on Domain Generalization Target Classification Method Based on Context Invariance Improvement

传统目标分类方法常常以测试数据集与模型训练数据集遵从一致数据分布的假设为前提,导致模型常产生被虚假关联特征误导分类结果、泛化性能差等问题。该研究利用上下文一致性构建具有良好上下文平衡效果的上下文估计器,对训练集进行重加权消除不平衡上下文对虚假关联特征的关注,进而提出了针对ERM基线模型融合Fish算法中基于梯度调整的元优化方法。实验表明,在PACS数据集上,该方法的准确率与基线模型相比提高了2.8%,相较当前最优主流领域泛化目标分类模型Fish提高了2.1%。

Traditional target classification methods often assume that the test dataset and the model training dataset follow a consistent data distribution,leading to problems such as the model being misled by false association features and poor generalization performance.This paper utilizes context invariance to construct a context estimator with good context balancing effect,reweights the training set to eliminate the attention of imbalanced context to false association features,and proposes a meta optimization method based on gradient adjustment for the ERM baseline model fusion Fish algorithm.The experiment shows that on the PACS dataset,the accuracy of our method has improved by 2.8%compared to the baseline model,and by 2.1%compared to the current optimal mainstream domain generalization target classification model Fish.