SoftPatch is an unsupervised anomaly detection method designed to handle noisy data in image sensory anomaly detection. The method introduces a memory-based approach that efficiently denoises data at the patch level. Noise discriminators are used to generate outlier scores for patch-level noise elimination before coreset construction. These scores are stored in a memory bank to soften the anomaly detection boundary. Compared to existing methods, SoftPatch maintains strong modeling ability of normal data and alleviates overconfidence in coreset. Comprehensive experiments on MVTecAD and BTAD benchmarks show that SoftPatch outperforms state-of-the-art methods in noisy scenarios and is comparable in noise-free settings. The paper highlights the importance of studying noisy data in unsupervised anomaly detection, especially in industrial applications where defects are hard to collect. Existing methods often assume clean training data, leading to performance degradation in noisy environments. SoftPatch addresses this by proposing a patch-level denoising strategy that improves data usage efficiency and enhances model robustness through noise discriminators. The method uses three noise discriminators to re-weight coreset examples, improving the model's ability to handle noisy data. Experiments on MVTecAD and BTAD datasets demonstrate that SoftPatch achieves superior performance in noisy scenarios. The method outperforms existing methods in terms of anomaly detection accuracy and robustness. The results show that SoftPatch is effective in both noisy and noise-free settings, making it a promising approach for practical applications. The paper also discusses the limitations of existing methods and the importance of addressing noisy data in unsupervised anomaly detection. The proposed method provides a new baseline for unsupervised anomaly detection with noisy data, offering a new perspective for further research.SoftPatch is an unsupervised anomaly detection method designed to handle noisy data in image sensory anomaly detection. The method introduces a memory-based approach that efficiently denoises data at the patch level. Noise discriminators are used to generate outlier scores for patch-level noise elimination before coreset construction. These scores are stored in a memory bank to soften the anomaly detection boundary. Compared to existing methods, SoftPatch maintains strong modeling ability of normal data and alleviates overconfidence in coreset. Comprehensive experiments on MVTecAD and BTAD benchmarks show that SoftPatch outperforms state-of-the-art methods in noisy scenarios and is comparable in noise-free settings. The paper highlights the importance of studying noisy data in unsupervised anomaly detection, especially in industrial applications where defects are hard to collect. Existing methods often assume clean training data, leading to performance degradation in noisy environments. SoftPatch addresses this by proposing a patch-level denoising strategy that improves data usage efficiency and enhances model robustness through noise discriminators. The method uses three noise discriminators to re-weight coreset examples, improving the model's ability to handle noisy data. Experiments on MVTecAD and BTAD datasets demonstrate that SoftPatch achieves superior performance in noisy scenarios. The method outperforms existing methods in terms of anomaly detection accuracy and robustness. The results show that SoftPatch is effective in both noisy and noise-free settings, making it a promising approach for practical applications. The paper also discusses the limitations of existing methods and the importance of addressing noisy data in unsupervised anomaly detection. The proposed method provides a new baseline for unsupervised anomaly detection with noisy data, offering a new perspective for further research.