X-ray security screening plays a pivotal role in maintaining safety at airports, railway stations, and other transportation hubs. This technology provides a reliable method to inspect the contents of luggage and other items quickly and non-intrusively. By leveraging intelligent algorithms, modern X-ray imaging technologies significantly enhance the effectiveness of security screening procedures. We experimented with multiple state-of-the-art deep learning object detection architectures based on convolutional neural networks (CNNs) and transformers in order to analyze X-ray baggage images. This research delves further into the use of transformer-based algorithms (DINO and Deformable DETR) for the identification of forbidden objects by varying the number of heads. The comparison includes a few CNN-based object detection algorithms, including YOLO-F and RTMDet. The best performance is achieved by the DINO model with 0.804 mAP for the CLCxray dataset and by Deformable DETR with 0.641 mAP for the PIDray dataset. The results demonstrate that varying the number of heads in Transformer-based algorithms significantly affects their performance in identifying forbidden objects in X-ray security images.