基于改进YOLOv7红外海上船舶检测算法  

An Infrared Maritime Ship Detection Algorithm Based onImproved YOLOv7

作　　者：饶兴昌 郑盈盈 陆万浩 黄孙港 RAO Xingchang;ZHENG Yingying;LU Wanhao;HUANG Sungang(School of Electronics and Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210000,China;School of Electronics Information Engineering,Wuxi University,Wuxi 214000,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,南京210000 [2]无锡学院电子信息工程学院,江苏无锡214000

出　　处：《电光与控制》2025年第4期23-30,共8页Electronics Optics & Control

基　　金：江苏省教育厅项目(22KJB140016)。

摘　　要：针对红外海上船舶图像检测在近岸密集、远海小目标以及低分辨率等场景下存在的误检、漏检等问题,并为了使模型更轻量化,提出了一种基于改进YOLOv7的红外海上船舶检测算法。为增强主干网络对船舶目标的特征提取能力,重新构造REP-DSConv-ELAN模块,替换原网络中的ELAN模块;其次在颈部网络中引入InceptionNeXt模块,减少因网络深度增加而造成船舶高维特征信息的丢失,更好地进行多尺度融合以提高船舶的检测效果;最后在检测头部分使用最小点距离的边界框回归损失函数MPDIoU,增强在低分辨率小目标场景下的检测能力。在红外船舶数据集上的实验表明:改进算法的精确率、召回率、平均精度均值较原YOLOv7算法分别提高了3.99、2.55和3.40个百分点,参数量由37.23×10^(6)降至31.98×10^(6)。综上,改进算法在保证红外船舶检测精度的同时能有效改善误检和漏检等问题。Aiming at the problems of false detection and missed detection in the infrared maritime ship image detection in the scene of near-shore dense,far-sea small target and low-resolution,as well as making the model lighter,an improved infrared maritime ship detection algorithm based on YOLOv7 is proposed.In order to enhance the feature extraction capability of the backbone network,REP-DSConv-ELAN module is reconstructed to replace ELAN module in original network.Secondly,the InceptionNeXt module is introduced into the neck network to reduce the loss of high-dimensional characteristic information caused by the increase of network depth,and to better carry out multi-scale fusion to improve the detection effect of ships.Finally,the boundary box regression loss function with minimum point distance,namely MPDIoU is used in the detection head to enhance the detection ability in the low-resolution small target scenes.Experimental results on infrared ship dataset show that the precision,recall and mean average precision of the improved algorithm are increased by 3.99,2.55 and 3.40 percentage points respectively,compared with original YOLOv7 algorith m,and parameters is reduced from 37.23×10^(6) to 31.98×10^(6).In conclusion,the improved algorithm can effectively ameliorate the problems of false detection and missed detection while ensuring the accuracy of infrared ship detection.

关 键 词：船舶检测 YOLOv7网络 动态蛇形卷积 InceptionNeXt模块 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]