**6. Conclusion**

This work explores the possibility of generating synthetically composite X-ray security imagery for training of CNN architecture to bypass the collecting a large amount of hand-annotated real-world X-ray baggage imagery. We synthesise highquality synthetically composited X-ray images using TIP approach and we present an extensive comparison on how real and synthetic X-ray security imagery affects the performance of CNN architecture for prohibited object detection in cluttered X-ray baggage images. Our experimental comparison demonstrates Faster R-CNN achieves the highest performance with mAP: 0.88 when trained on *Real* data (the good), followed by *Real+Synthetic* (the bad) and *Synthetic* (the ugly) over a three-class, {*Firearms, Firearm parts, Knives*}, prohibited item detection problem. This demonstrates a strong insight into the benefits of using real X-ray training data, also challenge and promise of using synthetic X-ray imagery.

In our future work, it is worth further investigating how to improve the effectiveness of synthetically composited imagery for training CNN architecture. Based on other work [20], a potential direction is to generate more diverse prohibited items images using generative adversarial networks (GAN). The generated prohibited item images then could be used for generating synthetic baggage images using TIP or similar.
