Image Anomaly Detection Using Geometric Relationship Knowledge Graphs

The patent introduces knowledge graphs as an intermediary component that mediates between limited anomalous data and the anomaly detection task. The knowledge graph stores pre-defined normal geometric relationships and contextual knowledge about object interactions, serving as a bridge that enhances detection accuracy without requiring large amounts of anomalous training data. This external knowledge source compensates for the scarcity of abnormal examples by providing rules for what constitutes normal behavior.

The system performs preliminary action by pre-populating the knowledge graph with normal geometric relationships and contextual knowledge before the actual anomaly detection process. This advance preparation of knowledge structures allows the system to efficiently compare observed geometric relationships against established norms, improving detection accuracy while minimizing the need for extensive anomalous data during training.

The patent segments the anomaly detection system into distinct functional modules: geometric relationship extraction from images, knowledge graph storage and querying, expert rule evaluation, and anomaly scoring. This segmentation allows each component to be developed and optimized independently, managing overall system complexity while achieving high detection accuracy through coordinated operation of specialized subsystems.

The knowledge graph serves multiple functions within the system: it stores normal geometric relationships, provides contextual knowledge about object interactions, and acts as a reference for evaluating anomalies. This multi-functionality reduces the need for separate specialized components, thereby managing complexity while maintaining high detection accuracy across different anomaly types.

The patent applies partial action by computing likelihood values selectively for the most relevant object pairs based on their geometric proximity and semantic relationships, rather than exhaustively evaluating all possible pairs. This approach maintains detection reliability by focusing computational resources on critical relationships while reducing overall computational time through intelligent sampling and prioritization.

The system employs self-service mechanisms by using the scene graph structure to automatically identify and prioritize object pairs that are most likely to contain anomalies based on their spatial and semantic relationships. This self-organizing approach reduces the need for external computational intervention and optimizes the evaluation sequence to minimize total processing time while maintaining reliability.