摘要
针对YOLO V3算法中对于小目标检测精度不高、容易出现漏检误检的问题,提出了一种基于改进YOLO V3的舰船目标检测算法。首先,通过在YOLO V3原网络结构基础上额外从主干网络引出一个输出尺度,将其与上一个输出尺度中的特征信息进行特征拼接,得到具有更丰富语义信息的特征向量;其次,基于数据集进行聚类改进,改进度量距离公式、重新设置anchor box的个数与相应参数;最后,优化改进YOLO V3的损失函数,提高模型的整体性能。对测试数据集进行分析实验,结果表明改进后的检测算法平均精确度达到83.98%,较之于原YOLO V3,平均精确度提升了6.72%。
The original YOLO V3 algorithm has low accuracy in small target detection and is prone to missed detection and false detection.To solve the problema ship target detection algorithm based on the improved YOLO V3 is proposed.Firstlybased on the structure of the original YOLO V3 networkan additional output scale is derived from the backbone networkwhose feature information is spliced with that of the prior output scaleso as to obtain a feature vector with richer semantic information.Secondlybased on the data setthe clustering is improvedin which process the distance measurement formula is improvedand the number of anchor boxes and the corresponding parameters are reset.Finallythe loss function of the improved YOLO V3 is optimizedso as to improve the overall performance of the model.Analysis and experimental results on the test data set show that the average detection accuracy of the improved algorithm is 83.98%which is 6.72%higher than that of the original YOLO V3.
作者
姜文志
李炳臻
顾佼佼
刘克
JIANG Wenzhi;LI Bingzhen;GU Jiaojiao;LIU Ke(Coast Guard Academy Naval Aviation University Yantai 264000,China;No.95668 Unit of PLA Kunming 650000,China)
出处
《电光与控制》
CSCD
北大核心
2021年第6期52-56,67,共6页
Electronics Optics & Control