Perpustakaan Digital ITB

This thesis develops and evaluates an AI-based object-detection approach to support airport baggage screening on dual-view X-ray imagery. Five prohibited-item classes: guns, hammers, wrenches, pliers, and bats, were annotated on a curated dataset and used to compare two deployment-ready detectors available on the Roboflow platform: a CNN-style model (Roboflow 3.0) and a transformer-family model (RF-DETR). After preprocessing and augmentation, 2,405 images were used for training, with separate validation and test splits. Performance was measured using mAP@50, precision, recall, and F1. Roboflow 3.0 achieved higher recall and F1 (96.0% and 96.7%) than RF-DETR (91.2% and 94.2%) while maintaining similarly high precision (>97%). Roboflow 3.0 also operated effectively at a lower decision threshold, which is desirable for safety-critical screening that prioritizes sensitivity to minimize missed threats. The results indicate that modern, off-the-shelf detectors, when tuned for high recall and paired with appropriate post-processing, can enhance consistency in checkpoint operations, while limitations such as class coverage and multi-view reasoning remain important directions for future work.