In-Su Han, Myung-Il Roh, In-Chang Yeo, Seong-Won Choi, Dohyun Chun, "A Method of Exemplar-Based Symbol Detection for Enhancing the Accuracy and Efficiency of Ship Fire and Safety Plan Review Processes", Proceedings of G-NAOE 2026, Houston, USA, 2026.10.20-24
| Abstract | In the ship design stage, a fire and safety plan is prepared to ensure safe operation by schematically depicting the safety equipment and evacuation routes required for firefighting, evacuation, and rescue. The fire and safety plan must comply with regulations such as the International Convention for the Safety of Life at Sea (SOLAS) and applicable classification society rules, which prescribe the required quantities and designated locations of safety equipment. Therefore, once the fire and safety plan is finalized, designers and surveyors typically perform manual checks to identify and count these symbols while reviewing the drawing for potential omissions. Recently, deep learning-based symbol detection has been explored in research to enhance the accuracy and efficiency of the review process. However, the same equipment is often drawn in different symbol styles across shipyards or individual designers. This variability hinders consistent model training and can cause detection performance to drop sharply when the model encounters unseen styles. To address this limitation, we proposed an exemplar-based symbol detection method that leverages the symbol table included in each safety plan. The table contains exemplar symbols that reflect the exact visual style used in the corresponding drawing, and we used these exemplars as model references during detection. By referencing the style-consistent symbol examples, the proposed method could maintain robust performance across drawings produced by different shipyards and designers, despite substantial variations in symbol style. We applied the proposed method to safety plans with diverse symbol styles not included in the model’s training data and evaluated its effectiveness by detecting symbols with high precision and recall. |
|---|---|
| Publication Date | 2026-10-20 |
