A robust multi-scale feature fusion model for low-quality multi-source data in underwater environments

Efficient and accurate perception of complex underwater scenes is crucial for ensuring the success of subsequent tasks. Multi-source image fusion techniques offer an effective solution, however, the presence of complex factors such as feature distortion, imaging blur, and lighting variations in low-quality multi-source (sonar-optical) underwater images leads to significant degradation in fusion performance. To address this issue, we propose a novel underwater multi-source data fusion model, incorporating multi-scale features detection and fusion. First, we extract shallow and deep features from multi-source data to detect rich local texture features and global structural features. Then, the detailed features and semantic information in the fusion process were enhanced through the designed multi-scale feature fusion module, and the problems such as low saturation and partial feature loss in the fusion image reconstruction were alleviated. This provides accurate multi-source fusion capabilities for downstream tasks. Extensive experiments on the public dataset demonstrate that our fusion method significantly improves the performance of tasks by 0.74% and 3.34%, surpassing related state-of-the-art methods.