Electronic Theses and Dissertation

Keamanan di fasilitas publik memerlukan sistem deteksi otomatis yang akurat dan efisien untuk mengidentifikasi barang terlarang pada citra X-ray bagasi. Penelitian ini melakukan analisis komparatif antara dua arsitektur deteksi objek terbaru, yaitu YOLO26s dan RT-DETRv4-S, yang keduanya menawarkan inferensi end-to-end tanpa Non-Maximum Suppression (NMS). Tiga dataset publik citra X-ray keamanan digunakan, yaitu PIXray (5.046 citra, 15 kelas), CLCXray (9.565 citra, 12 kelas), dan OPIXray (8.885 citra, 5 kelas). Kedua model dilatih menggunakan bobot pre-trained COCO dan dilakukan fine-tuning selama 50 epoch dengan augmentasi data dan penanganan ketidakseimbangan kelas melalui pembobotan frekuensi terbalik dan copy-paste augmentation. Hasil evaluasi menunjukkan RT-DETRv4-S secara konsisten mengungguli YOLO26s pada metrik mAP@50 di ketiga dataset dengan nilai 0,964 (PIXray), 0,877 (OPIXray), dan 0,701 (CLCXray), serta Recall tertinggi mencapai 0,9706 pada dataset PIXray, yang penting untuk meminimalkan risiko objek berbahaya yang lolos dari deteksi. YOLO26s mengungguli dalam efisiensi komputasi dengan throughput 60,7–70,1 FPS dan ukuran model 20,3 MB, dibandingkan RT-DETRv4-S sebesar 34,1–40,0 FPS dengan ukuran 42,2 MB. RT-DETRv4-S direkomendasikan untuk skenario yang mengutamakan akurasi deteksi tinggi, sedangkan YOLO26s lebih sesuai untuk sistem dengan keterbatasan sumber daya komputasi.

Security at public facilities requires accurate and efficient automated detection systems to identify prohibited items in X-ray baggage images. This research conducts a comparative analysis between two state-of-the-art object detection architectures, YOLO26s and RT-DETRv4-S, both offering end-to-end inference without Non-Maximum Suppression (NMS). Three public X-ray security datasets were used: PIXray (5,046 images, 15 classes), CLCXray (9,565 images, 12 classes), and OPIXray (8,885 images, 5 classes). Both models were trained using COCO pre-trained weights and fine-tuned for 50 epochs with data augmentation and class imbalance handling through inverse frequency weighting and copy-paste augmentation. Evaluation results show that RT-DETRv4-S consistently outperforms YOLO26s in mAP@50 across all three datasets with values of 0.964 (PIXray), 0.877 (OPIXray), and 0.701 (CLCXray), and the highest Recall reaching 0.9706 on the PIXray dataset, which is crucial for minimizing the risk of dangerous objects missed during detection. YOLO26s outperforms in computational efficiency with throughput of 60.7–70.1 FPS and model size of 20.3 MB, compared to RT-DETRv4-S at 34.1–40.0 FPS with 42.2 MB. RT-DETRv4-S is recommended for scenarios prioritizing high detection accuracy, while YOLO26s is more suitable for systems with limited computational resources.