基于YOLOv3的船只检测算法研究

Research on Ship Detection Algorithm based on YOLOv3

针对长江流域错综复杂的生态环境以及执法部门人员短缺对长江10年禁渔令实施的限制情况;通过智能视频监控系统对长江流域过往船只目标检测,对于判别船只有无违法捕捞行为具有重要意义。当前,传统的目标检测算法早已被检测效率更高、算法复杂度更低的深度学习方法替代;基于智能视频监控对于实时性的要求,采用YOLOv3作为目标检测模型,在兼顾检测精度的同时检测速度也更高。YOLOv3算法中,先验框作为目标检测算法的重要机制,影响着预测框的定位性能。在K-means聚类算法上进行改进,通过改变K值初始化随机选择不能获取全局最优解的情况,对K值选择时应用轮盘法,选择距离已经形成的聚类中心尽可能远的值作为新的K值,使各个聚类中心相对距离尽可能大,从而尽可能获得全局最优的聚类结果。实验结果表明,K-means优化后获得的先验框训练模型让船只目标检测性能更加优异,整体mAP提升了9.31%。

In view of the intricate ecological environment of the Yangtze River Basin and the limitation of the shortage of law enforcement personnel on the implementation of the ten-year fishing ban in the Yangtze River;in this paper,the intelligent video surveillance system is used to detect the target of passing ships in the Yangtze River Basin,which is of great significance for judging whether there is illegal fishing behavior.At present,traditional target detection algorithms have long been replaced by deep learning methods with higher detection efficiency and lower algorithm complexity.This paper considers the real-time requirements of intelligent video surveillance,and adopts the YOLOv3 algorithm under comprehensive consideration,which takes into account the detection accuracy and the detection speed is also higher.In the YOLOv3 algorithm,the prior frame is an important mechanism of the target detection algorithm,which affects the positioning performance of the prediction frame.In this paper,the K-means clustering algorithm is improved.By changing the K value initialization,the random selection the global cannot obtain the optimal solution,the roulette method is applied to the selection of the K value,and the value that is as far as possible from the cluster center that has been formed is selected as the new K value,so that the relative distance of each cluster center is as large as possible,so as to obtain the globally optimal cluster as much as possible.The experimental results show that the K-means optimized prior frame training model makes the ship target detection performance more excellent,and the overall mAP is increased by 9.31%.