Boosting-Inspired sequential multi-head learning for underwater acoustic target recognition

AbstractUnderwater acoustic target recognition (UATR) from ship-radiated noise is challenging due to strong intra-class variability and deployment-relevant domain shifts across targets and recordings. To improve robustness without computation burden or relying on extra annotations, we propose a boosting-inspired sequential multi-head learning framework built upon model-agnostic. Starting from a normally single-head baseline suffers from low diversity representation, we attach multiple lightweight classification heads to different depths of the network and train them sequentially. Each newly added head performs residual correction through an additive logit update, while a step-wise freezing protocol stabilizes optimization by locking previously trained heads. This design encourages a coarse-to-fine refinement of decisions and exploits multi-level acoustic representations in a principled manner. Experiments follow a recording-level split protocol to prevent clip leakage across train/test partitions. On ShipsEar, the proposed method improves ResNet18 accuracy from 63.96% to up to 70.31%, and increases the averaged score (mean of accuracy, precision, recall, and F1) from 61.36% to up to 68.75%. Gains consistently generalize across multiple encoders, and statistical tests confirm the improvements are significant. Overall, the proposed sequential residual multi-head strategy provides a simple, model-agnostic enhancement for practical UATR systems.