基于改进YOLOv5s的红外船舶目标检测算法

An Infrared Ship Target Detection Algorithm Based on Improved YOLOv5s

针对传统船舶检测算法参数量多、计算量大的问题,提出一种结合知识蒸馏(KD)的轻量级红外船舶目标检测算法CB-YOLOv5。在YOLOv5s主干网络引入轻量级网络Ghost模型以实现检测主干网络的轻量化;引用新的颈部结构渐进型特征金字塔网络(AFPN),AFPN通过融合两个相邻的低级特征,并渐进向高级特征融合,避免了非相邻级别较大的语义差距;使用VFL函数,改善红外船舶目标检测任务中正负样本不平衡的问题,以提升模型的整体性能;结合KD算法将学习能力强的教师网络中的“知识”迁移到所改进的网络模型中,来提升分类和定位的准确率。实验结果表明,在红外船舶数据集中,与原算法YOLOv5s相比,参数量减少了38%,mAP提升了3.9个百分点,模型权重文件大小仅为8.96×10^(6),证明了所提算法的有效性,具有一定的实用价值。

A lightweight infrared ship detection algorithm CB-YOLOv5 combined with Knowledge Distillation(KD)is proposed to solve the problem of large amount of parameters and computation in traditional ship detection algorithms.Lightweight network Ghost module is introduced in YOLOv5s backbone network to realize lightweight detection network.A new neck structure of Asymptotic Feature Pyramid Network(AFPN)is introduced,which can avoid the large semantic gap of non-adjacent level by fusing two adjacent low-level features and gradually fusing to higher-level features.The VFL function is used to improve the imbalance of positive and negative samples in infrared ship target detection tasks,so as to improve the overall performance of the model.Finally,KD is adopted to transfer the“knowledge”in the network of teachers with strong learning ability into the improved network model to improve the accuracy of classification and localization.Experimental results show that in infrared ship dataset,in comparison with original algorithm YOLOv5s,parameter amount is reduced by 38%,and mAP is increased by 3.9 percentage points,while the model weight file is only 8.96×10^(6),which proves the proposed algorithm is effective and has certain practical value.