Abstract:To address the constraints of detection capability, real-time performance, and computational resources in rapid obstacle identification for large-scale underactuated unmanned underwater vehicles (UUVs), a high-recall, rapid recognition method for small targets was investigated. A fast small-target recognition model based on an improved YOLO architecture is proposed: GhostNet is introduced as the backbone network, reducing the computational load by 48.75% through feature map reuse and linear transformation strategies. The Convolutional Block Attention Module (CBAM) is embedded in the neck network to enhance the feature response for small targets. Additionally, a dynamic weighted Focal-EIoU loss function is designed to alleviate the imbalance between positive and negative samples. Ablation and comparative experiments were conducted on a self-constructed sonar image dataset of fishing nets. The improved model""s performance metrics were as follows: the AP50 score reached 97.5%, an increase of 1.3 percentage points over the baseline model; the mAP improved by 5.7 percentage points to 58.4%; and the recall rate increased by 3.2 percentage points to 97.3%. The comprehensive performance surpasses that of mainstream models such as YOLOv8n and YOLOv11n. The triple synergistic optimization effectively balances detection speed and accuracy for underwater sonar images, significantly reduces the miss-detection rate for small underwater targets, and improves the robustness of small underwater target recognition. This research provides reliable technical support for real-time obstacle avoidance in large-scale underactuated UUVs.
