Classifier-free guided diffusion for side-scan sonar: Joint data enhancement with attention mechanism

Side Scan Sonar (SSS) is crucial for underwater environmental perception and indispensable for applications such as underwater target detection and topographic mapping. Nevertheless, the inherent complexity of underwater acoustic propagation, coupled with the substantial expense of acquiring high-fidelity imagery, results in a critical shortage of SSS datasets. This scarcity restricts the deployment of artificial intelligence (AI)-based solutions in autonomous underwater perception systems. To address this challenge, this paper implements explicit classifier-free guided generation of SSS images based on a diffusion model. First, we derive a controllable categorical generation model by linearly extrapolating between unconditional and conditional generative models. We utilize a guidance scale coefficient to govern the fidelity of category-specific generation. Furthermore, we effectively integrate categorical information and temporal information into the backbone network and introduce an attention mechanism to enhance the extraction of key shadow features. We also optimize the sampling speed. Finally, we validate the effectiveness of the proposed algorithm through comprehensive quantitative and qualitative experiments on a self-collected SSS dataset comprising real-world lake and marine environments. The results demonstrate the effectiveness of the proposed method.