FrYOLO: Fractional-Order Feature Propagation for Object Detection in Forward-Looking Sonar

Underwater object detection using forward-looking sonar presents fundamental challenges absent from terrestrial imagery: low-contrast single-channel inputs, multi-scale acoustic shadows, and object classes spanning a wide range of acoustic scattering characteristics. Three coordinated modifications to the YOLOv8 framework are proposed to address structural limitations of standard bottleneck chains for this domain. A fractional-order feature propagation mechanism based on Grunwald–Letnikov discretization enables each bottleneck to access a decaying-weighted history of all prior intra-chain feature states via a single learnable scalar per block. A boundary-aware gating module with joint spatial-channel attention selectively suppresses fractional correction at geometric boundary locations. A parameter-free energy-based attention module applied in the detection neck exploits the local statistical distinctiveness of genuine acoustic features during multi-scale fusion. Evaluated on the Underwater Acoustic Target Detection dataset, the proposed system achieves mAP50 of 0.8635 and mAP50-95 of 0.3964, improvements of 0.0188 and 0.0136 respectively over the YOLOv8n baseline at less than 2.0% parameter overhead, surpassing larger generic YOLOv8 variants on mAP50.