基于部分卷积的轻量化航拍船舶检测算法  

Lightweight Aerial Ship Detection Algorithm Based on Partial Convolution

作　　者：吴天翊 潘胜达[1] 安博文[1] WU Tianyi;PAN Shengda;AN Bowen(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《船舶工程》2025年第2期129-137,共9页Ship Engineering

基　　金：国家自然科学基金(62271302);上海市自然科学基金(20ZR1423500)。

摘　　要：[目的]针对无人机航拍船舶检测中存在的尺度多样性、低层次特征不足及设备性能限制问题,提出一种基于部分卷积的轻量化Bar YOLO算法。[方法]该算法在首层C2f结构中整合了BMA多尺度注意力模块,提升了对大型船舶的检测精度。通过并行的Par Bot模块替换主干网络的其余C2f模块,提升了特征提取效率。在颈部网络的C2f中采用部分卷积(PConv)技术,实现了模型的轻量化,并使用WIo Uv3加速模型收敛。[结果]在自制数据集M100ship上,改进算法的P_(A,m,50)和P_(A,m,50~95)分别提升了2.3%和0.9%,模型计算复杂度(GFLOPS)从8.1降至7.2,帧率(FPS)从91.8增至99.8,且在其他数据集上也展现出良好的泛化能力。[结论]研究结果表明,该算法可为边缘计算设备航拍船舶检测作业提供一定参考。[Purpose]To address the challenges of scale diversity,insufficient low-level features,and device performance limitations in unmanned aerial vehicle(UAV)ship detection,a lightweight BarYOLO algorithm based on partial convolution is proposed.[Method]The algorithm integrates a broadened efficient attention(BMA)multi-scale attention module within the first-layer CSP bottleneck with 2convolutions(C2f)structure,enhancing detection accuracy for large ships.The rest of the backbone network's C2f modules are replaced with parallel ParBot modules to optimize feature extraction efficiency.partial convolution(PConv)techniques are used in the neck network's C2f modules,achieving model lightweighting,and WIoUv3 is employed to accelerate model convergence.[Result]On the custom dataset M100ship,the improved algorithm increases P_(A,m,50)and P_(A,m,50~95)by 2.3%and 0.9%respectively,reduces the computational complexity(GFLOPS)from 8.1 to 7.2,and increases the frame rate(FPS)from 91.8 to 99.8.The algorithm also demonstrates good generalization ability on other datasets.[Conclusion]This algorithm provides a reference for edge computing devices in aerial ship detection operations.

关 键 词：船舶检测 注意力机制 部分卷积 

分 类 号：U671.99[交通运输工程—船舶及航道工程]