Research on Small Target Detection in Sonar Images Using Improved DEIM Based on Diffusion-GAN Data Augmentation

Abstract

Small-target detection in side-scan sonar images is challenging due to weak target responses, low contrast, complex target–shadow structures, and strong seabed texture interference, especially under limited annotated data. To address these challenges, this paper proposes a data–model co-optimization framework that combines Background-Anchored Local Natural Fusion Augmentation via Diffusion-GAN with an improved DEIMv2-S detector. At the data level, local statistical adaptation, same-image recovery, and cross-background fusion are introduced to guide target-conditioned diffusion generation with boundary-aware adversarial refinement. By preserving real sonar backgrounds and editing only the target and boundary regions, the proposed augmentation method produces more natural synthetic samples while reducing background distribution drift. At the model level, HR-STA, a Dual-domain Encoder, and Spectral FiLM are incorporated to enhance high-resolution spatial details and spatial–frequency representations. Experiments on a self-built lake side-scan sonar dataset show that the improved DEIMv2-S achieves 76.55% mAP@50 and 41.88% mAP@75, outperforming the baseline by 4.33 and 3.69 percentage points. With the proposed augmentation, performance further increases to 79.36% mAP@50 and 43.91% mAP@75, demonstrating the effectiveness of the framework for limited-sample sonar small-target detection.