一种跨级特征自适应融合的轻量化水声目标检测模型

A Lightweight Underwater Acoustic Target Detection Model With Cross-level Feature Adaptive Fusion

摘要:

针对声纳图像中高噪声、特征丢失和低分辨率导致的检测困难，研究提出了一种自适应跨阶段轻量级目标检测模型。方法上，设计轻量级特征提取网络并引入多尺度注意力模块以增强特征聚焦；利用焦点调制网络替代传统SPPF结构，提高关键区域的定位与识别；在预测阶段引入改进的自适应空间特征融合模块，在保持轻量化的同时扩展感受野并强化物理与语义特征表征。实验结果表明，该模型在实测声纳数据集上较现有方法参数量更小，检测精度提升约4.7%，速度提高超过18%。研究表明，该模型在声纳图像检测任务中实现了精度、速度与规模的良好平衡，为未来水下目标检测模型的设计提供了有效途径。

Abstract:

This study proposes an adaptive cross-stage lightweight detection model to address the challenges of high noise, feature loss, and low resolution in sonar image target detection. A lightweight feature extraction network with multi-scale attention was designed to enhance feature focusing, while a focus modulation network replaced the conventional SPPF to improve localization and recognition of key regions. Furthermore, an improved adaptive spatial feature fusion module was introduced in the prediction stage to expand the receptive field and strengthen both physical and semantic representations while maintaining compactness. Experimental results on real sonar datasets show that the proposed model achieves fewer parameters, about 4.7% higher detection accuracy, and over 18% faster speed compared with existing methods. The findings indicate that the model achieves a favorable balance among accuracy, efficiency, and complexity, offering an effective approach for future designs of underwater acoustic target detection.