基于YOLOv3及迁移学习的河道船舶目标检测算法

A River Ship Target Detection Algorithm Based on YOLOv3 and Transfer Learning

关键词:  YOLOv3  迁移学习  散料船舶  目标检测  采砂监管

Key words:  YOLOv3  transfer learning  granular material ship  target detection  sand mining supervision

摘要:

为加强对河道监控视频图像中散体物料采运船舶的监测和跟踪，从而辅助实现智能、高效的河道采砂监管和散体物料调度，基于You Only Look Once version 3（YOLOv3）算法及迁移学习提出一种河道散体物料船舶目标检测算法。首先使用COCO数据集训练初始的YOLOv3算法，得到模型的预训练权重；然后对从广西重要河道周围监控设备采集的采砂运砂船舶影像数据进行图像处理，得到高质量船舶数据集；最后以此数据集为驱动，利用迁移学习得到的预训练权重来训练针对河道采砂船等重点目标的YOLOv3检测模型。该模型采用Darknet-53作为主干网络，并融合了多尺度的特征图，从而实现对小、中、大等各类目标的检测。实验结果表明：该算法在测试集上的平均精度和检测速度分别达到98.00%和17.78 fps，对提高河道采砂监管效能和实现散体物料智能调度具有现实意义。

Abstract:

In order to strengthen the monitoring and tracking of granular material transportation ships in river monitoring video images, so as to assist in the realization of intelligent and efficient river sand mining supervision and granular material scheduling, a river granular material ship target detection algorithm is proposed based on You Only Look Once version 3 (YOLOv3) algorithm and transfer learning.Firstly, the COCO dataset is used to train the initial YOLOv3 algorithm, and the pre-training weight of the model is obtained.Then, the image data of sand mining and sand transportation ships collected from the monitoring equipment around the important rivers in Guangxi are processed to obtain a high-quality ship dataset.Finally, driven by this dataset, the transfer learning pre-training weight is used to train the YOLOv3 detection model for key targets such as river sand mining ships.The model uses Darknet-53 as the backbone network and integrates multi-scale feature maps to achieve the detection of small, medium and large targets.The experimental results show that the average accuracy value and detection speed of the algorithm on the test set reach 98.00% and 17.78 fps, respectively, which is of practical significance for improving the supervision efficiency of river sand mining and realizing the intelligent scheduling of granular materials.