基于细节增强的级联多分类光电船舶检测  

Cascaded multi-classification photoelectric ship detection based on detail enhancement

作　　者：徐志京[1] 谢安东 XU Zhijing;XIE Andong(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：《光电子．激光》2023年第3期241-249,共9页Journal of Optoelectronics·Laser

基　　金：国家重点研发计划(2019YFB1600605);上海市扬帆计划项目(20YF1416700)资助项目。

摘　　要：为提升无人船航行环境中船舶目标的感知精度,提出一种基于细节增强的级联多分类船舶检测模型。首先,提出混合平移数据增强(pan-mixed data augmentation,PMDA)算法,减少模型对船舶整体轮廓的依赖;其次,设计可变卷积平衡特征金字塔(deformable convolution-balanced feature pyramid,DC-BFP),提高模型对船体细节特征的提取能力;再次,将全连接层和卷积层联合构成级联交叉检测器(cascaded cross detector,CCD),提高模型对船体细节特征的解析能力;最后,采用标签平滑正则化(label smoothing regularization,LSR)方法,改善多分类检测的过拟合问题。在自建11分类光电船舶检测数据集MCSD11上进行消融和对比实验,特征提取结果和实验结果数据可视化表明,模型的各个改进部分能够提升船舶检测效果,平均精度达到了91.53%,相比主流的检测模型,算法得到大幅提升。In order to improve the perception accuracy of unmanned ships while sailing,a cascade multiclass ship detection model based on detail enhancement is proposed.First,apan-mixed data augmentation(PMDA)algorithm is proposed to reduce the network′s dependence on the overall outline of the ship.Secondly,a deformable convolution-balanced feature pyramid(DC-BFP)is designed to improve the model's ability to extract detail features of the hull.Thirdly,combine the fully connected layer and the convolutional layer to form a cascaded cross detector(CCD)to improve the model's ability to analyze the detail features of the hull.Finally,the label smoothing regularization(LSR)is used to improve the overfitting problem of multi-class detection.Ablation and comparison experiments are carried out on the self-built 11-category photoelectric ship detection dataset MCSD11.The feature extraction results and the data visualization of the experimental results show that the improved parts of the model can improve the ship detection effect,with an average accuracy of 91.53%,which is higher than that of the mainstream ones.The detection model algorithm has been greatly improved.

关 键 词：光电船舶检测 可变卷积平衡特征金字塔(DC-BFP) MCSD11数据集 细节增强 混合平移数据增强算法(PMDA) 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]