Advanced modeling for fire risk evaluation and intelligent preventive system
Abstract
Shipboard fires are among the most catastrophic maritime incidents, as dense smoke, fluctuating lighting, and camera movement often obstruct timely detection. To address these operational challenges, we propose a lightweight yet robust vision-based fire and smoke detection system tailored for vessels with limited onboard computing resources. The proposed model integrates a partially fine-tuned ResNet-34 feature extractor with a capsule-based detection head, which preserves spatial relationships critical for identifying irregular flame and smoke patterns. Directional attention mechanisms are employed to highlight subtle, low-contrast smoke formations, which frequently occur under maritime conditions such as haze or sea glare. When evaluated on a newly created marine dataset representing a variety of real-world scenarios, the model demonstrates exceptional accuracy and real-time inference performance on edge-class devices. Our solution, which incorporates the fine-tuned ResNet-34 backbone, a PAA module, and a capsule-based detection head, outperforms all comparative methods, achieving a maximum [email protected] of 90.17%, an IoU of 72.58%, Precision of 92.2%, and Recall of 88.5%. These results indicate that the model excels at both precise localization and detecting subtle or partially obscured fire and smoke regions—an ongoing challenge in shipboard imaging.