Label noise is widely present and unavoidable, and it affects the performance of deep network models. Sample selection methods based on the principle of small loss can easily and effectively handle label noise by the “memory effect” of neural networks. This study proposes a new sample selection principle and a two-stage weighted sample selection and relabeling method (WSSR-2s) based on the principle that a closer sample distance in the feature space results in more similarity, combined with the assumption of high and low confidence of the samples. In the early training stage, for high-confidence samples, their voting rights are weighted in the feature space to better guide training. In the middle and later stages of training, for low-confidence samples, their voting rights are transferred to their most similar feature samples for more accurate training. The experimental results on synthetic noise datasets CIFAR-10 and CIFAR-100, as well as real noise datasets ANIMAL-10N and WebVision, show that the proposed method achieves higher accuracy and can better handle label noise problems.