To address the performance degradation of deep neural networks caused by noisy labels, we propose hard sample adaptive labeling with optimal reweighting for noisy labels, a novel hard sample adaptive weighting method. By reweighting and retraining the model multiple times in each epoch, learning from different subsets of hard samples, and iteratively predicting pseudo-labels, HEALON improves the accuracy of noisy label correction. Experiments demonstrate that our method outperforms existing approaches on noisy label learning tasks, showing significant performance gains and better generalization. This research provides a new perspective for tackling the label noise problem in real-world scenarios.