A Dual-Branch Vision–Sonar Fusion Detector for Underwater Robotic Object Detection

Abstract

Underwater robotic perception requires robust object detection under severe optical degradation and acoustic ambiguity. Optical images provide rich appearance and boundary cues but are sensitive to turbidity, scattering, and illumination changes, whereas sonar images provide stable structural responses but contain sparse texture, low resolution, acoustic shadows, and artifacts. This article proposes a YOLOv8-style dual-branch vision-sonar fusion detector for underwater robotic object detection. Paired optical and sonar images are processed by two homologous but parameter-independent backbones to extract P3, P4, and P5 features. At each scale, a bidirectional gated cross-modal fusion block performs sonar-guided visual enhancement and vision-guided sonar enhancement, and produces enhanced visual, enhanced sonar, and fused feature streams. Three parallel PAN necks and decoupled detection heads are supervised jointly, with the fused branch used as the primary output. Experiments on a paired UMOD vision-sonar subset show that the fused branch achieves AP50=0.8485, recall = 0.8936, and mAP@0.5:0.95=0.4416 on the test set. It improves AP50 over the sonar branch and independent single-modal baselines, and obtains higher AP50 and recall than an AquaYOLO-style sonar-only reproduction under the same protocol. Qualitative and failure-case analyses show that fusion improves target recovery but may degrade under severe modal conflict or weak visual cues.