Improved YOLOv9 for Detecting Weapons: Using Transfer Learning, Multi-Scale Training, and Ensemble Techniques to Achieve Higher Accuracy

The detection of accurate weapons will enhance society with the correct assistance; that is why surveillance systems exist to maintain people's safety. In this study, weapon detection is improved through the YOLOv9 framework, which includes notable improvements in training and inference techniques. The transfer learning approach leveraging a pretrained YOLOv9 model was adopted, where the YOLOv9 pretrained model was further trained to achieve the fine-tuned model on the domain-specific weapon dataset that resolved problems, including slow training and poor generalization of the features by the model. Multi-scale training was performed to improve model robustness to both weapon size and image quality for real-life applications. Finally, ensemble methods were used to combine predictions from YOLOv9 and YOLOv8 to harness their complementary strengths to improve detection accuracy and reliability. These improvements together show the possibility for better performance in weapon detection use cases, thus future-proofing themselves for reliable real-time use in security and surveillance systems. In conclusion, the improvements boosted the main metrics like mAP, recall, accuracy, and precision, and these results show their fairness.