This paper proposes a Double Correction Framework for Denoising Recommendation (DCF), which addresses the challenges of noisy samples in recommendation systems. The framework consists of two components: sample dropping correction and progressive label correction. The sample dropping correction component aims to accurately identify and retain hard samples while dropping noisy samples. It uses confirmed loss calculation and cautious hard sample search to stabilize loss values and retain informative samples. The progressive label correction component iteratively re-labels highly determined noisy samples and re-trains them to further improve performance. The framework is evaluated on three datasets and four backbones, demonstrating its effectiveness and generalization. The results show that DCF outperforms existing methods in terms of performance and generalization. The framework mitigates the adverse effects of noisy interactions by improving the stability of model predictions and enhancing the accuracy of label correction. The key contributions of this work include analyzing the limitations of loss-based dropping methods, designing two correction components to enhance denoising recommendations, and experimentally validating the effectiveness of the proposed framework. The framework is implemented in Python and is available at https://github.com/bruno686/DCF.This paper proposes a Double Correction Framework for Denoising Recommendation (DCF), which addresses the challenges of noisy samples in recommendation systems. The framework consists of two components: sample dropping correction and progressive label correction. The sample dropping correction component aims to accurately identify and retain hard samples while dropping noisy samples. It uses confirmed loss calculation and cautious hard sample search to stabilize loss values and retain informative samples. The progressive label correction component iteratively re-labels highly determined noisy samples and re-trains them to further improve performance. The framework is evaluated on three datasets and four backbones, demonstrating its effectiveness and generalization. The results show that DCF outperforms existing methods in terms of performance and generalization. The framework mitigates the adverse effects of noisy interactions by improving the stability of model predictions and enhancing the accuracy of label correction. The key contributions of this work include analyzing the limitations of loss-based dropping methods, designing two correction components to enhance denoising recommendations, and experimentally validating the effectiveness of the proposed framework. The framework is implemented in Python and is available at https://github.com/bruno686/DCF.