Explainable Multitype Anomaly Detection Using a Lightweight Fully Convolutional Architecture

Explainable anomaly detection methods can identify and localize anomalies within images but typically cannot distinguish anomaly types. Furthermore, they often require separate models for each object category, increasing training and maintenance costs. Recent vision‐language models address this problem but remain computationally expensive. To address this, we propose Multitype Fully Convolutional Data Description (MultiTypeFCDD), a unified, lightweight convolutional approach to explainable anomaly detection that trains on image‐level labels yet predicts type‐specific spatial heatmaps across multiple objects from a single model. Our primary contribution is a novel multichannel model, discriminative loss and training framework, enabling decoupled representations of co‐occurring defects while predicting anomaly maps in a single forward pass. We evaluated the method on the Real Industrial Anomaly Detection (Real‐IAD) manufacturing dataset, achieving 96.4% image‐level Area Under the Receiver Operating Characteristic (AUROC) curve, and the Water Research Centre (WRc) sewer defect dataset, achieving 92.2% image‐level AUROC. The model uses just over 1% of the parameter count of state‐of‐the‐art vision‐language models used for similar tasks and achieves a tenfold speed‐up over Gradient‐Weighted Class Activation Mapping (Grad‐CAM) (9.4 ms versus 94.7 ms) by avoiding computationally expensive backward passes. These results demonstrate the efficiency and adaptability of our framework across diverse industrial domains.