Accurate and real-time recognition of harmful objects in surveillance images to improve the security of public spaces

Accurate and real-time recognition of harmful objects such as firearms and cold weapons in surveillance images plays a vital role in improving the security of public spaces. However, existing models often suffer from performance degradation when faced with real-world challenges such as sudden changes in lighting, small target sizes, and objects occlusion. In this paper, an efficient hybrid approach based on the YOLOV۵s model is presented for recognizing harmful objects in the challenging HOD dataset. The proposed approach is built upon two main basis: (۱) designing a class-specific data augmentation pipeline in which four categories of optical, geometric, structural, and environmental transformations are applied in accordance with the visual characteristics of each class, and (۲) integrating the Efficient Channel Attention (ECA) module into the YOLOv۵s architecture, which enhances the ability to extract distinctive features without noticeably increasing model complexity. Experimental results show that the proposed YOLOV۵s+ECA model achieves a mean average precision of mAP۵۰ of ۸۵.۲۰% on the class-specific augmented HOD dataset, representing an improvement of ۲.۸% over the YOLOV۵s baseline. This gain in recognition accuracy is achieved with the addition of only three parameters to the proposed model and a slight reduction in speed from ۱۲۰ to ۱۱۰ FPS (frame per second). Ablation studies further confirm that the two innovations improve recognition performance through complementary pathways: class-specific data augmentation raises the recall rate by expanding training diversity, while the ECA module reduces false positives and enhances precision by focusing on key features.

کلیدواژه ها:

Real-time recognition ، Harmful objects ، Class-specific data augmentation ، HOD dataset ، YOLOv۵s+ECA model