A Sequence-Aware Framework for Maritime Signal Flag Recognition and Semantic Interpretation

Automatic recognition of maritime signal flags requires not only identifying individual flags but also reconstructing their ordered combinations and interpreting the corresponding maritime meanings. This study develops a sequence-aware framework that integrates visual perception, ordered sequence reconstruction, reliability validation, and International Code of Signals (ICS)-based semantic interpretation. A maritime signal flag dataset containing 7865 images from 40 ICS categories was collected using a single small model vessel in sunny weather. The visual perception module employs Flag-YOLO, an enhanced YOLO11-based detector integrating DualConv, GSConv, and Dynamic Head to improve the recognition of small, distant, blurred, and wind-deformed flags. Detected flags are subsequently arranged in top-to-bottom order, validated using confidence and spatial-consistency constraints, and interpreted using a 313-entry ICS subset restricted to single-group hoists of up to three flags without substitute pennants. Flag-YOLO achieves 93.1% mAP@0.5 and 61.9% mAP@0.5:0.95 with 2.1 M parameters. In a sequence-level evaluation of 1500 samples, the proposed approach achieves 92.9% accuracy, compared with 88.5% for YOLO11n, with the largest improvement of 7.2 percentage points observed for three-flag sequences. The results demonstrate the effectiveness of integrating visual perception, structured sequence recognition, and domain knowledge for automated maritime visual communication understanding, providing a potential perception component for intelligent maritime systems.