Toward Collaborative Multimodal Information Perception: A Hierarchical Feature Alignment Network and Acoustic-Optical Dataset for Underwater Object Detection

In the realm of Earth observation, underwater target detection serves as a critical component for characterizing aquatic environments. However, existing acoustic-optical collaborative methods face persistent challenges, including the high calibration difficulty across different modalities and unsatisfactory feature alignment. A core difficulty lies in enabling the network to adaptively capture intrinsic dependencies and achieve efficient fusion of multimodal information during the learning process. To address this, this paper proposes an object detection method with adaptive cross-modal feature alignment. Specifically, the method establishes automatic alignment relationships between sonar and optical images through fine-grained and instance level feature fusion. Furthermore, a novel fusion strategy is designed to enable efficient acoustic-optical information fusion via same-layer feature learning. Finally, to mitigate the issues of scarcity and low quality of multimodal underwater datasets, a data acquisition platform is constructed, and a strictly spatiotemporally synchronized sonar-optical dataset is built. Extensive experimental results demonstrate that the proposed method simultaneously enhances detection performance on both sonar and optical images. On the created dataset, the proposed method achieves mAP@0.5 scores of 86.9% and 94.4% for acoustic optical collaborative detection, thereby outperforming main stream approaches. In addition, the robustness and generalization of the proposed method are verified in other multimodal object detection scenarios. The proposed AOMU dataset can be found at https://github.com/LaplaceXing/GDOU/tree/master.