STMH-Xception: Sequential multi-head residual learning for robust underwater acoustic target recognition

Underwater Acoustic Target Recognition (UATR) remains a formidable challenge due to highly non-stationary marine environments and the complex, overlapping acoustic signatures of different vessels. Deep Neural Networks (DNNs) typically employ a single-head optimization paradigm, rendering them susceptible to overfitting individual-vessel noise rather than learning robust, category-consistent features. To bridge this gap, we propose a novel framework, the Sequentially Trained Multi-Head Xception (STMH-Xception). Drawing inspiration from gradient boosting, we introduce a step-wise STMH strategy that integrates deeply supervised auxiliary heads into a depthwise separable convolutional backbone. Unlike standard joint-training multi-task learning, STMH employs a sequential freezing protocol whereby each subsequent head targets and corrects the residual representation errors of its predecessors. This "coarse-to-fine" learning scheme drives the network to extract a highly discriminative feature hierarchy while mitigating inter-head gradient conflicts associated with joint multi-head optimization. Validated on the ShipsEar and DeepShip datasets, the proposed STMH-Xception increases the classification accuracy from 68.72% to 73.10% and from 72.89% to 76.14%, respectively, compared to the baseline. Supported by statistical significance tests and generalizability analyses across diverse CNN encoders, the STMH strategy emerges as a highly efficient, CNN-backbone-agnostic optimization paradigm. The additional parameter and computational cost introduced by the auxiliary heads remain small, making the framework promising for resource-constrained UATR scenarios.