UniSLAD: A Unified Framework for Structural and Logical Industrial Visual Anomaly Detection

UniSLAD is a unified framework for detecting both structural and logical anomalies in industrial automation, addressing a gap in existing methods. It employs a dual-feature extractor combining a CNN for local texture and a Transformer for global context, alongside advanced feature representation techniques such as memory banks with the Mahalanobis Transform and aggregation methods like Lower-Upper Mean and Power Mean Pooling. The framework demonstrates competitive performance on industrial benchmarks, achieving 99.4% and 93.1% accuracy, making it a valuable tool for practitioners needing robust anomaly detection in dynamic environments.