Improving object detection efficiency using backspacing loss

The paper considers the problem of automatic processing of aviation security screening images. Special attention is paid to methods of classification of prohibited items on X-ray images of luggage and hand luggage. It is proposed to use modified training loss functions. It is shown that this approach provides a gain of 2-4% over traditional training. Moreover, a cascade object detection model based on YOLOv8 neural network and binary classifier is developed. Studies have shown that the application of an additional classifier improves the performance of YOLO neural network by 2-3%. It should also be noted that this study used an in-house dataset prepared jointly with the Ulyanovsk Institute of Civil Aviation. Performance analysis was also performed for different versions of the YOLO detector, which showed that the best combination of accuracy and speed is provided by the medium-sized model (YOLOv8m). The main result of the work is to improve the efficiency of detection and classification of prohibited items by using a modified loss function and cascade detector.