摘要
舰船作为主要的海上交通作战工具,在遥感图像中高效准确识别舰船目标具有重要研究意义。光学遥感舰船图像包含丰富的信息,但因其具有复杂度高、图像大、受天气和昼夜变化影响等特点,导致识别率较低。针对这一问题,本文通过对ResNet18进行改进,提出一种更加高效的光学遥感舰船图像分类的方法。对ResNet18网络进行了简化,降低其参数量;使用并行池化实现特征图的空间降维,在保持特征丢失较少的情况下加快网络收敛;引入多尺度卷积进行不同尺度特征信息的提取,并使用ECA注意力机制改进多尺度卷积模块与残差模块,解决分支网络支路融合时存在特征不能很好的在通道间交互的问题。在FGSCR-42数据集上进行实验,实验结果表明改进后的算法收敛速度更快,且准确率与F1-score均高达95%左右,较ResNet18网络提高了7%左右,而参数量仅有改进前的20%左右;与其他网络在舰船目标识别中的性能相比,本文方法也更加出色。
As the main means of Marine traffic warfare,it is of great significance to identify ship targets efficiently and accurately in remote sensing images.Although the optical remote sensing ship image contains rich information,the recognition rate is low because of its high complexity,large image and the influence of weather and day and night change.To solve this problem,this paper proposes a more efficient optical remote sensing ship image classification method by improving ResNet18.The ResNet18 network is simplified and its parameter number is reduced.The parallel Pooling is used to reduce the dimensionality of the feature graph space to speed up the network convergence while keeping less feature loss.The multi-scale convolution is introduced to extract feature information of different scales,and the ECA attention mechanism is used to improve the multi-scale convolution module and residual module to solve the problem that features can′t interact well between channels in branch network branch fusion.Experiments were carried out on the FGSCR-42 dataset,and the experimental results show that the improved algorithm converges faster,and the accuracy and F1-score are up to about 95%,which is about 7%higher than that of the ResNet18 network,while the number of parameters is only about 20%of that before the improvement.Compared with the performance of other networks in ship target recognition,the proposed method also has better performance.
作者
曾富强
张贞凯
方梦瑶
Zeng Fuqiang;Zhang Zhenkai;Fang Mengyao(Ocean College,Jiangsu University of Science and Technology,Zhenjiang 212003,China)
出处
《电子测量技术》
北大核心
2024年第12期164-172,共9页
Electronic Measurement Technology
基金
国家自然科学基金面上项目(61871203)资助。
作者简介
曾富强,硕士研究生,主要研究方向为计算机视觉、深度学习。E-mail:221112201127@stu.just.edu.cn;通信作者:张贞凯,教授,博士,主要研究方向为雷达信号处理、深度学习等。E-mail:zhangzhenkai@just.edu.cn。
