Image Classification with Noisy and Unlabeled Samples
Jinwu He, Yipeng Yan · 2024
In deep learning, data annotation poses a significant challenge, often resulting in datasets with missing labels or noisy annotations, both of which can severely degrade model performance. Addressing these issues is crucial for enhancing the accuracy and generalization of models. In this framework, multiple teacher models collaborate to produce robust pseudo-labels, while multiple student models utilize regularization terms to correct noisy annotations. Specifically, for labeled samples with noisy annotations, the teacher models generate more reliable labels from multiple perspectives by aggregating prediction results, thus alleviating the noise issue. For unlabeled samples, the label with the highest confidence from the teacher models’ outputs is selected as the pseudo-label. Experimental results demonstrate that this approach significantly improves model performance when handling datasets with both noisy and missing labels, highlighting its potential for deployment in practical applications.