MLAD: Manifest and latent anomaly detection based on the integration of reconstruction and MLFP-KNN methods

Abstract In the realm of industrial manufacturing, anomaly detection technologies are essential for enhancing production efficiency. This paper aims to optimize the detection of anomalies in industrial manufacturing settings by introducing a novel framework manifest and latent anomaly detection (MLAD). Firstly, we redefine anomalies as manifest and latent, further subdividing the original categories of structural and logical anomalies. To address the issue of insufficient feature space information utilization in previous reconstruction-based methods, which led to inadequate detection of latent anomalies, we combine a reconstruction-based approach with the multi-layer feature pyramid k-nearest neighbors technique. Additionally, we introduce a Hybrid Attention mechanism to refine the autoencoder’s reconstruction process, thereby enhancing the detection of latent anomalies and increasing overall accuracy. Our experiments on the MVTec LOCO AD dataset demonstrate that MLAD achieves a significant improvement in anomaly detection, with an Image-level AUROC score of 90.8% and Pixel-level AU-sPRO of 79.3%. The integration of the Hybrid Attention mechanism significantly boosts the precision of anomaly detection, highlighting the effectiveness of our approach in enhancing industrial anomaly detection capabilities. The framework provides a more nuanced understanding of industrial defects, contributing valuable insights and practical improvements to the field of industrial anomaly detection.