基于双流注意力机制引导和多尺度特征表达的轻量化水下目标检测算法

Lightweight Underwater Target Detection Algorithm Driven by Dual-stream Attention Mechanism and Multi-scale Feature Representation

摘要:

在复杂的自然环境中,提升水下生物资源探测效率对中国海洋经济发展具有重要意义。为了应对水下目标检测中计算资源受限、水下环境复杂性所导致的检测效果差的问题,提出了一种基于改进YOLO v8s的水下目标检测算法CBM-YOLO。首先,设计了轻量化特征提取模块CSP-DLN,该模块融合了3×3和1×1卷积核的优势,能够高效捕捉细节空间特征,减少冗余计算。其次,为解决水下生物目标特征信息丢失的问题,引入新的特征融合网络BGL-FPN,通过跨尺度连接结合全局和局部空间注意力机制,有效提高检测精度。最后提出最大池化下采样(MPD),通过最大池化和卷积分支的并行处理,更好地捕捉小目标的边缘和细节,从而增强小目标的检测能力。实验结果表明,该算法在URPC2020数据集和UDD数据集上的mAP@0.5分别达到78.2%和69.1%,相较基准模型mAP@0.5各自提高1.5、2.6个百分点,同时参数量和计算量分别下降47.3%和25.5%,相比于其他主流目标检测算法表现性能也最优。该模型被部署到JetsonTX2嵌入式端并经过TensorRT加速推理后,mAP@0.5为77.4%,检测速度达到37.6f/s,能够在保证高检测精度的同时实现水下生物的实时检测。

Abstract:

In complex natural environments, enhancing the detection efficiency of underwater biological resources is vital for China??s marine economic development. To address the issues of limited computational resources and poor detection due to underwater complexity, CBM-YOLO, an improved YOLO v8s-based underwater target detection algorithm was proposed. Firstly, CSP-DLN, a lightweight feature extraction module, was designed, combining the advantages of 3 × 3 and 1 × 1 convolutional kernels to efficiently capture detailed spatial features and reduce redundant calculations. Secondly, to address the loss of feature information in underwater biological targets, a new feature fusion network, BGL-FPN, was introduced, which effectively improved detection accuracy through cross-scale connections combined with global and local spatial attention mechanisms. Lastly, max pooling downsampling ( MPD) was proposed, leveraging parallel processing of max pooling and convolutional branches to better capture edges and details of small targets, thereby enhancing their detection capability. Experimental results indicated that the algorithm attained mAP @ 0. 5 of 78. 2% and 69. 1% on the URPC2020 and UDD datasets, respectively, with improvements of 1. 5 and 2. 6 percentage points over the baseline model, while reducing parameters and computations by 47. 3% and 25. 5% . It outperformed other mainstream target detection algorithms. Deployed on the Jetson TX2 embedded device and accelerated by TensorRT, the model achieved mAP@ 0. 5 of 77. 4% and a detection speed of 37. 6 f/ s, enabling real-time underwater detection with high accuracy.