基于改进ConvNet的轻量化SAR舰船识别  

Lightweight SAR ship recognition based on improved ConvNet

作　　者：李炫潮 何永华 朱卫纲 李永刚 李晨瑄 LI Xuanchao;HE Yonghua;ZHU Weigang;LI Yonggang;LI Chenxuan(Graduate School,Space Engineering University,Beijing 101416,China;Department of Electrical and Optical Engineering,Space Engineering University,Beijing 101416,China)

机构地区：[1]航天工程大学研究生院,北京101416 [2]航天工程大学电子与光学工程系,北京101416

出　　处：《航天工程大学学报》2024年第2期71-79,共9页

摘　　要：深度学习技术为合成孔径雷达(SAR)舰船识别提供了新方法,但传统卷积神经网络在捕获目标全局信息方面存在局限,影响识别性能。文章结合Transformer全局特征提取优势和ConvNet的局部特征提取能力,提出了一种改进的卷积运算模块——全局循环卷积,该模块通过全局卷积核和循环卷积机制,有效提取全局特征,同时保持了模型的高效性和轻量化。通过一维分解,将二维卷积分解为一维卷积,进一步降低了模型的计算复杂度和参数量。实验采用OpenSARShip-4数据集,对比了基于全局循环卷积模块的ResNet18、MobileNetV2和ConvNeXt模型与传统模型的性能。结果显示,提出的模型在保持较低计算成本的同时,显著提高了识别精度。该研究为SAR图像中的舰船识别提供了一种有效的轻量化解决方案,对于促进相关技术的发展具有重要价值。Deep learning technology has introduced a novel approach for ship detection using Synthetic Aperture Radar(SAR).However,traditional convolutional neural networks have limitations in capturing global target information,which affects recognition performance.To this end,a modified convolution operation module—Global Recirculating Convolution(GRC)is proposed in this paper,which combines the global feature extraction capabilities of the Transformer with the local feature extraction strengths of Convolutional Networks(ConvNet).This module effectively extracts global features through a combination of global convolutional kernels and recirculating convolution mechanisms,while maintaining the efficiency and lightweight nature of the model.By decomposing the two-dimensional convolution into one-dimensional convolution through a dimensionality reduction technique,the computational complexity and parameter count of the model are further reduced.Experiments conducted using the OpenSARShip-4 dataset compared the performance of ResNet18,MobileNetV2,and ConvNeXt models equipped with the Global Recirculating Convolution module against traditional models.The results demonstrate that the proposed models significantly enhance recognition accuracy while maintaining low computational costs.This research provides an effective lightweight solution for ship detection in SAR imagery,holding significant value for the advancement of related technologies.

关 键 词：SAR舰船识别 全局循环卷积 一维分解 轻量化 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]