基于改进VGG网络的单阶段船舶检测算法

单阶段多框架目标检测算法在目标检测领域取得 了成功的应用,但其针对公共数据集中船舶检测的平均精度明 显低于其它刚体类目标类别,同时现有公开数据集中的船舶数量较少且类别单一。为提高检 测精度,提出一种基于改 进VGG网络的单阶段船舶检测算法,在原有VGG底层网络的基础上加入异步卷积和最大池化的 交替连接结构,保证 实时处理的同时提高船舶检测的平均精度。为增加训练所需的船舶数量和类别,广泛收集互 联网中包含船舶的图片, 建立了包含22507个船舶目标的数据集,其中6902个目标标签细分为七类船舶。实验将公开数据集VOC2007和 VOC2012中的图片缩小至300训练后,SSS D在VOC2007test中的平均检测精度均值可达79.3%,平均检测速度 超过40 fps。通过迁移参数的方法,在自建数据集中训练后,对大类 船舶检测的平均精度超 过84%,对七类船舶检测的平均精度均值超过89%,领先现有同类船舶检测 算法。

Single stage ship detection algorithm based on improved VGG network

Single shot MultiBox detector has been successfully applied in the dom ain of target detection,while its average accuracy of ship in the public datasets is obviously lower than that of other rigid-body targets.Meanwhile,the number of ship in the public datasets existing is small and the category of ship is single.In order to impro ve the accuracy of detection,a single stage ship detector based on the improved VGG network was proposed.The alternating connection struc ture of asynchronous convolution and maximum pooling was added to the underlying network of VGG to ensure real-time detectio n and improve the average accuracy of targets detection.In order to increase the number and the types of ships needed for tra ining,images of ships were widely collected by means of internet skills,and a dataset containing 22507ships was e stablished,among which 6902ships were subdivided into seven types.After the training of VOC2007trainval and VOC2012trainval with Nominal Res olution of 300×300,the average precision of SSSD in VOC2007test can reach 79.3%,and the average speed of ship detections exceeds 40fps.After transforming the corresponding parameters and training on proposed datasets,the average accuracy of large clas s ships detection exceeds 84%,and that of seven class ships exceeds 89%,leading the existing similar algorithms for ships detection.