Detection of violence and threats from images using artificial intelligence methods

In this thesis, the problem of automatic detection of threat and violence elements from images is addressed. For this purpose, a manually annotated dataset consisting of four classes knife, weapon, violence, and non-violence was constructed. All annotations were prepared using bounding boxes and converted into the COCO format for experimental evaluation. Four deep learning–based object detection models, namely YOLOv8, Faster R-CNN, RetinaNet, and Mask R-CNN, were evaluated. To ensure the generalizability of the results, all models were trained and tested using a 10-fold cross-validation strategy. The training process was structured into three stages: early learning, balancing, and evaluation, and only the averaged results obtained during the evaluation stage were reported. Experimental results show that YOLOv8 achieved the highest precision of 98.3%, while its recall of 74.7% indicates that some threat instances were missed. Faster R-CNN provided the most balanced performance with a recall of 98.0% and a accuracy of 97.6%. Mask R-CNN produced comparable results, achieving 97.6% recall and 96.8% accuracy. In contrast, RetinaNet exhibited lower performance, with a recall of 46.1%. Overall, the results quantitatively demonstrate that two-stage region-based models provide more stable and reliable performance than single-stage detectors for threat and violence detection on the constructed dataset.