摘要
以海上目标的红外、可见光图像为数据源,针对海上目标尺度多样、数据源多波段信息丰富的特点,基于YOLOv3原型网络架构,根据FPN原理将4倍降采样获取的第11层底层特征图与第103层深层特征图进行融合,实现对网络尺度的扩展,并通过K-means聚类算法得到更为精细化尺度下的先验框;同时将红外、可见光图像根据目标特点按比例进行组合,形成图像源的物理层融合,进而构建混合数据集进行多波段协同模型训练。实验结果表明,S4-YOLO网络模型其识别的准确率高于YOLOv3、YOLOv3-Tiny模型,可以很好地适应海上多尺度目标的识别需求。
Taking the infrared and visible image of marine targets as data source,considering the characteristic of multiscale and multi-band information of data source for marine targets,based on the prototype YOLOv3 network architecture and FPN principle,the bottom feature map of the 11th layer is fused with the deep layer feature map of the 103rd layer to achieve the expansion of the network scale,and k-means clustering algorithm is used to obtain the prior box at a more refined scale.Meanwhile,the infrared and visible images are combined in a certain proportion to form the physical layer fusion of the image source,and then the mixed data set is constructed for multi-band cooperative model training.The experimental results show that the recognition accuracy of S4-YOLO network model is higher than that of YOLOv3 and YOLOv3-Tiny models,and it can adapt to the identification needs of multi-scale marine targets.
作者
赵文强
孙巍
ZHAO Wen-qiang;SUN Wei(Equipment Department of the Navy inWuhan Area Military Representative Office,Wuhan 430064,China)
出处
《光学与光电技术》
2020年第4期38-46,共9页
Optics & Optoelectronic Technology
关键词
海上目标
深度学习
卷积神经网络
检测识别
多尺度
marine targets
deep learning
convolution neural network
detection and recognition
multi-scale
作者简介
赵文强(1984—),男,硕士,工程师,主要研究方向为图像与信号处理。E-mail:465650366@qq.com