Various industries attempting to adopt artificial intelligence technology struggle to solve data scarcity due to security and privacy concerns. In the defense industry, tanks vary in appearance by country, and their design may change due to functional improvements. Additionally, collecting images of enemy tanks on the field is particularly challenging. Generative models used for data augmentation in recent studies sometimes require additional data. Our experiment explored the feasibility of data augmentation techniques based on primitive methods, such as combining masked backgrounds and object images. This approach allowed us to create a large amount of data more simply compared to data shared on open platforms. The accuracy of the trained model with transfer learning and our virtual dataset outperformed the model trained on open data by 5.1 %.