Shadow-Aware Region-of-Interest Guided Multi-Scale Deep Feature Framework for Automatic Target Recognition in Side-Scan Sonar Images

Automatic target recognition in side-scan sonar imagery remains a challenging task due to speckle noise, acoustic shadow distortion, limited contrast variation, and complex seabed interference patterns. Conventional texture-based feature extraction approaches demonstrate limited robustness when detecting small-scale underwater objects under heterogeneous seabed conditions. Recent deep feature learning architectures improve localization performance; however, detection accuracy is still affected by background clutter and weak highlight–shadow structures. A shadow-aware region-of-interest guided multi-scale detection framework is introduced for enhanced localization of mine-like underwater targets in side-scan sonar images. The proposed framework integrates speckle-resistant contrast enhancement, acoustic shadow-guided region extraction, and multi-resolution feature refinement prior to deep feature-based detection. Region-of-interest selection suppresses irrelevant seabed regions and improves discriminative representation of highlight–shadow target signatures. Multi-scale detection improves recognition of small and partially occluded objects frequently observed in sonar strips. Experimental evaluation conducted on publicly available side-scan sonar datasets demonstrates significant improvement in localization reliability compared with classical patch-based feature matching strategies and baseline deep detection configurations. The proposed framework achieves improved precision, recall stability, and detection robustness under complex seabed conditions, indicating suitability for automatic target recognition applications in underwater exploration, maritime safety monitoring, and mine-like object localization tasks.