Analytical residual network architecture with semi-supervision for sonar-based ship classification in underwater defense systems

Abstract

Sonar-based ship classification is vital for underwater defense systems, enabling effective surveillance, threat detection, and autonomous navigation. However, challenges such as high noise levels, low resolution, and complex acoustic scattering in sonar data necessitate the use of advanced algorithms. This study aims to develop a novel semi-supervised framework, Attention-ResNet, to enhance ship classification accuracy by integrating residual networks (ResNets) and attention mechanisms, thereby leveraging both labeled data (LD) and unlabeled data to address the scarcity of LD. The proposed Attention-ResNet framework combines ResNets with attention mechanisms to improve feature extraction and discriminative capability. It processes sonar signals as single-channel images, utilizing skip connections in ResNets to learn complex acoustic features and attention gates to focus on relevant signal regions. The framework is evaluated on two benchmark sonar datasets, DeepShip and ShipsEar, using semi-supervised learning with only 25% LD. Ablation studies assess the contributions of ResNet and attention components in both the image domain (ID) and audio domain (AD). The Attention-ResNet framework achieves a classification accuracy of 70.17% on the test dataset, a 10.59% improvement over the baseline. The alternative Attention-ResNet_2 architecture further improves accuracy to 71.92%, a 12.34% enhancement. Comprehensive ablation studies confirm the synergistic effect of ResNet and attention mechanisms in enhancing classification performance in both ID and AD. The Attention-ResNet framework demonstrates significant improvements in sonar-based ship classification, offering a robust solution for underwater surveillance and navigation systems. Its ability to leverage unlabeled data makes it particularly suitable for scenarios with limited LD. Future work will explore its application to diverse datasets and real-world implementations to enhance its practical utility further.