Spectro-Geometric YOLO: A Lightweight Framework for FLS Object Detection via Frequency Dynamics and Retentive Masking

Autonomous Underwater Vehicles (AUVs) rely on forward-looking sonar (FLS) for navigation in turbid waters. However, generic lightweight detectors often struggle to accommodate the specific acoustic characteristics of sonar imagery. Standard architectures frequently fail to differentiate target signatures from spectral noise and neglect the distance-dependent energy decay characteristic of underwater acoustics. To address these limitations, we propose SG-YOLO (Spectro-Geometric YOLO), a lightweight detector that integrates acoustic inductive biases into the detection pipeline. First, we incorporate enhanced residual frequency-dynamic (ERFD) blocks to separate semantic features from speckle noise within the spectral domain. Second, we introduce an adaptive channel-frequency retentive masking (ACF-RMT) mechanism to formalize highlight-shadow geometry and acoustic attenuation for clutter suppression. Finally, a slim-BiPAN neck is designed to prune redundant deep layers, aligning the receptive field with the small scale of sonar targets. Experimental results on the UATD dataset demonstrate that SG-YOLO achieves a balance between accuracy and efficiency, attaining 86.0% mAP₅₀ and 39.2% mAP_50:95 with only 0.63 million parameters and an inference speed of 428 FPS.