Abstarct: With the increasing deployment of visual surveillance systems in urban and public environments, automatic detection of security threats, particularly firearms and bladed weapons, has become a critical challenge in computer vision. Surveillance images and videos are often affected by low resolution, poor lighting conditions, noise, and compression, which significantly impact the performance of detection systems. Therefore, evaluating detection models under real-world conditions is of great importance. In this thesis, an automatic weapon detection system baxsed on deep learning is presented, and its performance is evaluated under different input data scenarios. The main focus of this study is to analyze the impact of input data quality and image preprocessing without modifying the standard structure of the detection model. Several scenarios, including raw input data, compressed images, and wavelet-baxsed preprocessing, are defined and evaluated in a fair and consistent experimental frxamework. Experimental results indicate that the baxseline scenario without preprocessing provides the best overall balance between precision, recall, and localization accuracy. Contrary to common assumptions, most preprocessing techniques do not lead to performance improvements and, in some cases, degrade detection accuracy. Qualitative analysis of image and video results further demonstrates that the proposed system shows reasonable stability in realistic surveillance conditions, although challenges such as small and partially occluded objects remain. The findings highlight the importance of realistic experimental evaluation in the design and deployment of practical weapon detection systems.