Možnosti detekce zbraní pomocí metod hlubokého učení

Abstract

This bachelor’s thesis addresses weapon detection using deep learning methods. The objective was to design, implement, and validate a convolutional neural network-based solution and to evaluate its accuracy. A reproducible pipeline was built, and selected YOLO variants were tested. Training was performed on a dataset containing images of pistols and knives with corresponding annotations. Model performance was assessed using mAP, precision, and recall. The results show accuracy differences across configurations, with more stable scores for medium-sized variants and for clearly visible targets; small and partially occluded objects remain challenging. Future research should focus on this issue, particularly in relation to the limited size and variability of the data. To achieve better results, it would be advisable to train the model on a significantly larger and more diverse dataset.