Real-Time Visual Anomaly Detection for Industrial Bottle Cap Inspection

Automated visual inspection is increasingly used to replace or support manual quality control in industrial production lines, where consistency, speed, and reliability are critical. This thesis presents the design, implementation, and experimental evaluation of a real-time anomaly detection system for the automated inspection of bottle caps on an industrial line, developed in collaboration with Giacobazzi winery. The system is designed to inspect four geometrically and optically distinct cap variants at production-line speed, without interrupting the inspection process when individual frames cannot be processed. The proposed pipeline combines a variant-specific geometric alignment stage, implemented through slot detection and template-matching strategies selected via a configurable rotation registry, with a PatchCore-based anomaly detection model. The model compares locally extracted patch features against a memory bank built from nominal training images. The software architecture separates global and cap-specific configuration through a two-tier design based on SQLite and YAML, allowing additional cap variants to be integrated without modifying the core inference pipeline. Several anomaly detection configurations were evaluated, including different backbone architectures and baseline methods. PatchCore achieved the best overall performance and was therefore selected for the final system. The resulting pipeline was evaluated in terms of both detection performance and inference latency, showing that the proposed memory-bank-based one-class approach can satisfy the accuracy and real-time requirements of the target industrial inspection task.