Label noise poses a challenging task, impacting the model’s generalization when trained with noisy labels. Existing noise is primarily categorized as symmetric/asymmetric noise, instance-dependent noise, and real-world noise. Most current methods are effective under specific assumptions, limiting their ability to handle all types of noise. To address this issue, we propose a debiased sample selection method. Initially, we ensure noise sparsity in each training iteration through sample correction. Subsequently, to achieve accurate sample partitioning, we employ implicit regularization and debiasing techniques to obtain more robust representations. Additionally, we use feature-based K-nearest neighbors (KNN) for clean sample selection during training. To enhance generalization, we incorporate common consistency regularization techniques. Extensive experiments demonstrate the effectiveness of our approach.