CBAM Anomaly Detection

Attention-guided autoencoder for one-class anomaly detection and localization

An attention-guided convolutional autoencoder for one-class anomaly detection and localization on CIFAR-10, combining CBAM (Convolutional Block Attention Module) with reconstruction-based anomaly scoring.

Training uses only normal samples; anomalies are identified through reconstruction error, while the attention maps provide spatial localization of where the anomaly occurs — an interpretability benefit that plain autoencoders lack.

Stack: PyTorch, CBAM, convolutional autoencoders

View on GitHub