Smart Security System Using Object Detection for Real-time Weapon Detection

This paper examined how the scope of fine-tuning in a YOLOv8-based weapon detection system shapes both detection quality and the interpretability of model decisions. Three configurations — frozen backbone, selective head and neck adaptation, and full-parameter optimization — were compared on a curated weapon surveillance dataset using complementary detection and explainability metrics. The key finding is that composite quality — the product of accuracy and spatial transparency — peaks at the intermediate fine-tuning level. Selectively adapting the detection head and neck raises [email protected] from 0.821 to 0.883 and improves WRAR on clear frames from 0.35 to 0.67, while preserving weapon-frame WRAR at 0.88. Fully optimizing all layers pushes mAP slightly higher to 0.901 but drives WRAR on weapon frames down to 0.19, with Grad-CAM and SHAP maps converging on the same conclusion: the model has substituted image-border cues for weapon appearance in its decision process. This collapse is statistically robust (p < 0.001; Cohen's d ≈ 1.18). Selective fine-tuning emerges as the deployment-ready choice: [email protected] of 0.883, weapon recall of 96%, the lowest missed-detection count, and spatially coherent explanations that withstand operational audit. These results position fine-tuning depth as a substantive architectural variable in security AI design — one that governs not just benchmark scores but whether a deployed system can be responsibly explained, validated, and trusted. The smart security system successfully demonstrated that integrating deep learning-based object detection with real-time surveillance infrastructure provides an efficient, accurate, and scalable solution.

[1]     D. Bhatt, C. Patel, H. Ghetia, and P. J. Bhatt, "Weapon detection using deep learning: A review," in Proc. IEEE ICCIT, Dhaka, Bangladesh, 2021, pp. 1–6.   [2]  R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, "Grad-CAM: Visual explanations from deep networks via gradient-based localization," in Proc. IEEE ICCV, Venice, Italy, 2017, pp. 618–626.   [3]  S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," in Adv. Neural Inf. Process. Syst., vol. 30, 2017, pp. 4765–4774.   [4]     G. Jocher, A. Chaurasia, and J. Qiu, "Ultralytics YOLOv8," Ultralytics, 2023. [Online]. Available: https://github.com/ultralytics/ultralytics.   [5]     C. Feichtenhofer, H. Fan, J. Malik, and K. He, "SlowFast networks for video recognition," in Proc. IEEE ICCV, Seoul, South Korea, 2019, pp. 6202–6211.   [6]  Z. Liu, J. Ning, Y. Cao, Y. Wei, Z. Zhang, S. Lin, and H. Hu, "Video Swin Transformer," in Proc. IEEE CVPR, New Orleans, LA, USA, 2022, pp. 3202–3211.   [7]   L. Oakden-Rayner, J. Dunnmon, G. Carneiro, and A.      Re, "Hidden stratification causes clinically meaningful failures in machine learning for medical imaging," in Proc. ACM Conf. Health Inference Learning, 2020, pp. 151–159.   [8]     A. S. Ross, M. C. Hughes, and F. Doshi-Velez, "Right for the right reasons: Training differentiable models by constraining their explanations," in Proc. IJCAI, Melbourne, Australia, 2017, pp. 2662–2670.   [9]   A. Saleem, J. Akhtar, and M. Usman, "Automatic weapon detection in surveillance videos: A survey," Sensors, vol. 21, no. 9, p. 2954, Apr. 2021. [10]  S. Akcay, A. Atapour-Abarghouei, and T. P. Breckon, "GANomaly: Semi-supervised anomaly detection via adversarial training," in Proc. ACCV, Perth, Australia, 2019, pp. 622–637.