Performance Enhancement of Drowning Detection Using Synthetic Data

초록

Drowning accidents remain a critical global safety issue, yet vision-based deep learning approaches are hindered by the scarcity of real drowning data due to ethical and legal constraints. This paper proposes a synthetic data-driven drowning detection framework that combines structured prompt-based image generation with a Residual-GAN to construct a balanced dataset of 8,400 synthetic images (400 prompt-generated seeds and 8,000 GAN-generated samples). A lightweight residual CNN (two Residual Blocks with Global Average Pooling) is then trained exclusively on synthetic data. To assess real-world applicability, we perform synthetic-to-real cross-domain validation using an independent test set collected from the Roboflow Universe public dataset platform (500 drowning and 500 non-drowning images) without fine-tuning. The proposed model achieves 98.0% validation accuracy and 94.3% test accuracy, demonstrating performance comparable to AlexNet (94.8%) and ResNet18 (94.9%) while maintaining a significantly more compact architecture. These results indicate that controllable synthetic data generation combined with lightweight residual learning can provide an effective foundation for drowning detection under real-data scarcity.

키워드