An Improved YOLOv12 Algorithm for Small Object Detection Based on the FLSMDD Dataset

Detecting small targets in forward-looking sonar imagery is indispensable for ocean engineering, marine archaeology, and autonomous underwater navigation, yet the task is hampered by the low resolution, heavy speckle noise, and size imbalance that typify FLS data; to mitigate these hurdles, we adapt YOLOv12 with three coordinated improvements—multi-scale feature fusion, a lightweight convolutional block, and a tailored augmentation pipeline—each tuned to the Forward-Looking Sonar Marine Debris Detection set so that the network can isolate subtle echoes more reliably. When benchmarked, our enhanced YOLOv12 surpasses conventional detectors on mean average precision, precision, and recall, all while sustaining real-time throughput, and ablation analysis attributes most of this gain to the fusion hierarchy and its accompanying attention layers, which amplify fine-grained details without inflating computation. These findings point to a practicable route for smarter sonar-based inspection of small underwater targets, with immediate relevance to engineering operations; looking ahead, we will pursue further slimming of the network, explore cross-modal fusion with optical or acoustic channels, and refine on-device deployment strategies for field-ready platforms.