Detecção de materiais ilícitos em imagens de raio-x de bagagens em aeroportos

Abstract

Air transport is growing significantly in Brazil and the flow of users is constantly expanding. Given this, ensure the safety of users who must accompany this growth. Among the essential security measures at airports is the airworthiness of baggage. This measure takes place during the boarding flow and intends to look for objects that may contain some risk during the users' flight. This task takes place through an X-ray in which the luggage is kept, generating a real-time image, used by human operators to carry out the analysis, and thus detect whether the luggage contains any object that could pose danger. However, this verification needs a lot of attention and training of the operators, so that they do not cause mistakes during the analysis of the objects presented in the image. Factors such as: skill, motivation, among others, can directly affect the analysis, any failure can expose flights to threats. During the dispatch procedure, some airports rely on the aid of integrated systems that carry out simple object classification. We propose a convolutional neural network capable of classifying the objects contained in the bags and even saying the possible objects that present a danger to the flight, with a confidence factor that can help human operators in a more significant way. During the methodology, it discusses a convolutional neural network model, based on the YOLO architecture, which aims to classify objects in real time. A search was carried out to find some X-ray baggage image banks. From the definition of the image banks, image processing techniques were applied, to better highlight the characteristics of the images, in addition to leaving the standardized images following the standards of the network used. It was noted the need to apply a data argumentation method that carried out the expansion of image banks. Two classification methods were proposed: binary and non-binary, where objects such as: knives, weapons, portable chargers, among others, as dangerous. Each method carried out the labeling of the images in a different way, the binary method was labeled using an area of the image that may or may not contain a dangerous object, while the non-binary model carried out the labeling specifically on objects that may be dangerous. From the finalization of the labeling, the training of the YOLO convolutional neural network used in the methodology was carried out. The two methods used in classifying the datasets had great results during the classification of objects contained in luggage, achieving an accuracy of 96% for the non-binary classification and 67% accuracy for the binary model.