【摘要】 为解决传统基于重构的异常检测方法重建图像质量低导致异常检测准确度不高的问题，提出了一种基于记忆增强扩散模型的异常检测方法。考虑到时间成本以及计算资源有限，使用潜在扩散模型作为基础架构。同时为更好的避免异常部分的直接重建，引入了记忆增强模块记住正常数据的典型特征，从而使异常数据的重构误差更大，提高了异常检测的准确性。为了在保证正常区域相同的情况下重建异常区域，提出了一种噪声条件嵌入的方法，提高了重建的稳定性。在MVTec-AD上的实验结果表明，与相关方法相比，所提方法有更好的检测和定位性能。更多还原

【Abstract】 To address the issue of low anomaly detection accuracy caused by poor image reconstruction quality in traditional reconstruction-based anomaly detection methods, an anomaly detection method based on a memory-enhanced diffusion model was proposed. Considering time costs and limited computational resources, the latent diffusion model was employed as the foundational architecture. Additionally, to better avoid the direct reconstruction of anomalous regions, a memory-enhanced module was introduced to memorize typical features of normal data, thereby increasing the reconstruction error for anomalous data and improving the accuracy of anomaly detection. To reconstruct anomalous regions while preserving normal regions, a noise condition embedding method was proposed, enhancing reconstruction stability. Experimental results on MVTec-AD show that the proposed method outperforms related approaches in both detection and localization performance.更多还原