Exemplar-based comparative learning for visual anomaly detection under high feature diversity and spatial entropy

Automated anomaly detection is essential for industrial inspection, yet conventional pattern-recognition approaches are limited by their reliance on predefined anomaly categories. To address this challenge, we propose ECHO (Enhanced Comparative model for High-entropy anOmalies), a deep neural network for unsupervised anomaly localization based on comparative inference. ECHO employs dual convolutional branches to extract features from baseline and query images, followed by a comparison network that identifies salient discrepancies indicative of anomalies. The proposed framework was evaluated across multiple application domains, including PCB assembly inspection, obstacle detection, missing-object inspection, marble surface inspection, and injection-molded part inspection. In addition, the public ADFI dataset was used to benchmark ECHO against state-of-the-art anomaly detection methods, including PatchCore and CutPaste. Experimental results demonstrate that ECHO achieves strong cross-domain generalization and consistently outperforms conventional approaches such as YOLOv8n, single-input CNNs, and template matching. Saliency map analysis further confirms that ECHO focuses on anomaly-relevant regions and provides interpretable decision-making. Across all case studies, ECHO achieved accuracies exceeding 80% and effectively detected anomalies with ambiguous normal–abnormal boundaries, where conventional object detection methods often failed. Furthermore, while ECHO exhibits strong transferability without application-specific training, targeted fine-tuning further improves detection performance.