| Title |
Improving Detection Performance via R-VAE-Based Synthetic Data Augmentation for Indoor Fire Scenarios |
| Authors |
김경민(Gyeongmin Kim) ; 고진환(Jinhwan Koh) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.8.84 |
| Keywords |
Fire detection; Generative AI; Synthetic data; Data augmentation; Variational autoencoder |
| Abstract |
This paper proposes an efficient synthetic data generation and augmentation methodology to address the severe data imbalance and scarcity issues associated with accident data, such as indoor fires. Unlike conventional approaches like VAE, GAN, or Diffusion models that generate entire images, we designed a 'Residual learning-based VAE (R-VAE)' architecture that leverages the background invariance of indoor CCTVs to separate the background and learn only the variations (residuals). The methodology consists of a two-stage pipeline: first, a commercial Text-to-Image (T2I) model is utilized to generate a small number of initial fire seed data by combining a fixed background with fire text prompts; second, these seeds are fed into the R-VAE for large-scale self-augmentation. This approach maximizes computational efficiency by eliminating unnecessary background noise interference. Furthermore, we introduced parallel encoders/decoders configured with 1x1, 3x3, and 5x5 kernels to capture the diversity of fire features. Evaluating the quality of the generated data, the proposed R-VAE model (based on the parallel structure) recorded an FID of 163.38 and a KID of 0.1136, demonstrating superior image generation quality compared to a conventional GAN model (FID 233.14, KID 0.2188). Notably, it also exhibited approximately four times greater training time efficiency than the GAN. To verify the practical utility of the augmented data, it was applied to train high-level classification models, including ResNet-50 and GoogLeNet. The results confirmed that Accuracy, Precision, and F1-Score improved significantly compared to using simple synthetic data alone. Finally, an analysis of performance variations by progressively increasing the augmentation ratio to 0.5, 2, 3, and 5 times the seed data revealed that the classification models' performance improved stably as the ratio increased. At an augmentation ratio of 2x or higher, most evaluation metrics achieved a high reliability score of over 0.99. |