Contrastive Adaptation on Domain Augmentation for Generalized Zero-Shot Side-Scan Sonar Image Classification

Predicting never-seen-before targets in underwater side-scan sonar (SSS) recognition is challenging due to limited data availability and complex environmental factors. Traditional supervised methods achieve high accuracies in standard tasks but fail to generalize in zero-shot scenarios. Recent style-based transfer-learning methods for zero-shot learning (ZSL) in SSS have shown promise but suffer from unrealistic environmental and sample assumptions. To overcome these limitations, we propose a contrastive adaptation of domain augmentation (CADA), a novel learning paradigm that utilizes background fusion and noise modeling to expand generalized ZSL (GZSL), offering greater practicality in engineering. By integrating simulated SSS noise with fused backgrounds, our approach augments unseen classes, improves class separability, and mitigates overfitting. The contrastive adaptation further narrows domain distribution gaps while preserving critical intraclass semantic content information. Moreover, we introduce the first SSS image dataset tailored for the GZSL application. Experimental results show that CADA reaches up to 73.32% on the harmonic mean index, achieving over 20% higher accuracy than existing state-of-the-art style-based methods, highlighting its effectiveness for SSS target classification in GZSL settings. The code is https://github.com/JiaYP0433/CADA-Generalized-Zero-Shot-Side-Scan-Sonar-Image-Classification.