基于轻量化YOLOv8n的船舶号灯识别  

Ship’s lights identification based on lightweight YOLOv8n

作　　者：乔亚静 高祥雨 赵月林[1] QIAO Yajing;GAO Xiangyu;ZHAO Yuelin(Navigation College,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]大连海事大学航海学院,辽宁大连116026

出　　处：《大连海事大学学报》2025年第1期82-91,共10页Journal of Dalian Maritime University

基　　金：辽宁省“兴辽英才计划”自主项目(XLYC1902071)。

摘　　要：为解决船舶号灯目标检测中参数计算量大、背景光线复杂及号灯种类繁多等问题,对YOLOv8n做出了改进,以满足实时准确识别船舶号灯的要求。首先,使用VanillaNet网络作为主干特征提取网络,以降低模型的计算成本,满足实时检测要求;其次,引入颜色注意力模块确定和增强号灯的颜色特征,以提高在复杂光线背景下号灯颜色的识别能力;再次,针对号灯的大小、频率、各视图空间排列等特征,构建MoE-layer模块代替C2f模块,以提高船舶号灯识别的准确性;最后,通过Focal Loss调节对难易分类样本的关注来解决类别不平衡问题,提高模型对号灯目标的检测能力。实验结果表明,相较于原基线YOLOv8n模型,改进后的模型在参数量和计算量分别降低37.7%和52.8%的情况下,精度和mAP@0.5分别提升3.3%和2.2%,达到98.3%和98.7%,可以满足实时准确识别船舶号灯的要求。To address the large number of parameter volume,intricate complex background lighting and the diversity of the ship’s lights in the context of target detection,improvements on the YOLOv8n were made to meet the requirements for real⁃time and accurate identification of ship’ s lights. Firstly, theVanillaNet network was used as the backbone feature extrac⁃tion network to reduce the calculation cost of the model.Sec⁃ondly,the color attention module designed was designed todiscern and amplify the distinctive color attributes of the navi⁃gation lights to improve the recognition of the lamp color in acomplex light background. Then, to accommodate the uniquecharacteristics of the ship’s lights, a Mixture of Experts(MoE)⁃layer module was built as a substitute for the conven⁃tional C2f module to further refine the identification accuracy.Finally, the Focal Loss function was used to adjust the focuson easy⁃to⁃hard classified samples so as to improve the abilityof the model to detect the ship’s lights. The experimental re⁃sults show that compared with the original baseline modelYOLOv8n, the improved model has 37. 7% fewer parametersand 52.8% less computation, respectively, while the precisionand mAP@ 0.5 have increased by 3.3% and 2.2%, respec⁃tively, reaching to 98.3% and 98.7%. Hence, the improvedYOLOv8n can meet the requirements of real⁃time accurateidentification of ship’s lights.

关 键 词：船舶号灯识别 目标检测 VanillaNet MoE混合专家层模块 Focal Loss 

分 类 号：U675.3[交通运输工程—船舶及航道工程]