Adaptive feature refinement with information preservation for multiclass unsupervised anomaly detection

Recent advancements in anomaly detection have shown significant potential across various industrial domains. However, a wide range of unpredictable defect types emerge in the real world, and anomaly images are often challenging to obtain, making traditional methods less suitable. To address this challenge, recent studies have focused on the multiclass unsupervised anomaly detection task. Nonetheless, these approaches face significant challenges owing to the difficulty in robustly handling diverse classes and defect types. We propose adaptive feature refinement anomaly detection (AFRAD), which integrates a multilayer perceptron-based stage-adaptive decoder that adaptively decodes multiscale feature maps to model broader contextual relationships. Furthermore, to minimize information loss, we introduce a convolution neural network-based focused local decoder to capture fine details at low-level dimensions and an MLP-based compensation decoder. The compensation decoder compensates for information missed by the stage-adaptive decoder and focused local decoder. This strategy improves the ability of the model to handle diverse aspects of the data, enabling robust anomaly detection. In addition, we experimentally demonstrate that the fusion of final representations enables the generation of high-quality reconstructed feature maps. Our AFRAD achieves superior performance compared with conventional reconstruction-based methodologies on various public datasets.