CogniScan: cognitive AI-based unsupervised anomaly detection AI framework for X-ray luggage screening: integrating statistical patterns and contextual relationships

CogniScan is a novel Cognitive AI related framework which was built on an unsupervised anomaly detection approach in X-Ray luggage screening. The framework is designed to mimic human nature on the visual inspection process in X-Ray luggage screening and decision-making. The system deviates from traditional supervised learning methods that require labelled threat examples, and the proposed systems learn through unlabeled data by identifying anomalies based on statistical pattern recognition and contextual relationship analysis. The framework created on two complementary approaches: (1) GMM – Gaussian Mixture Models for the task of recognizing statistical patterns which demonstrate outperforms K-Means clustering across qualitative and quantitative analysis, and (2) GNN – Graph Neural Networks for contextual relationship analysis for which compare three significant approaches: one-class, binary-class, and unified training approach. Based on the comparative analysis, it revealed that even though one-class approach achieved marginally higher numerical performance, the unified GNN approach significantly overshadowed with good computational efficiency and deployment practicality with only modest performance trade-offs. Three ensemble approaches – OR, AND, and Combined Score—were considered for the decision fusion. After the comprehensive analysis it was revealed that Combined Score approach ended up with a balance between threat detection sensitivity and operational efficiency. Experiments on X-Ray luggage scans which were collected from residential security environment in Sri Lanka further demonstrated that CogniScan framework has succeeded with high threat detection (98% sensitivity) and good efficiency (62.95% specificity), demonstrating that it has a good threat detection capability by maintaining a practical operational efficiency, without the requirement of labeled data. This research has made a significant foundation for adaptable and transparent security screening systems which can handle varying threats.