Threat Object Classification in X-ray Images Using Transfer Learning

Abstract

Automatic classification of threat objects in X-ray images is important because of terrorist incidents happening in every country especially in the Philippines. Manual inspection using X-ray machine is prone to human error due limited amount of time given to the operator to check the baggage. This task is also stressful because there are lots of objects to be identified and needs full attention. Over long period of time, the performance of human inspector decreases and the chance of missed detection increases. As a solution to the problem, this paper used the concept of transfer learning in classification of threat objects. The threat objects used in the experiment consists of 4 classes such as blade, gun, knife and shuriken. The dataset came from the GDXray database, a public database of X-ray images. Experiment results showed that by using the concept of transfer learning with data augmentation and fine-tuning, threat objects can be classified at 99.5% accuracy.