魏松杰　张泽栋　徐臻　刘梅林　陈伟

摘   要：针对SAR图像中舰船目标的检测问题，单纯基于深度学习的图像处理技术难以达到检测准确性和实时性要求. SAR图像中目标尺寸较小，且易受噪声、光斑干扰，传统方法难以提取精細特征并克服复杂条件下的背景干扰. 针对以上问题，设计基于YOLOv3检测框架的端到端检测模型，借鉴了残差模块结构来避免网络退化问题. 同时结合深层与浅层的不同尺寸特征图检测，使用目标基础特征提取网络参数来避免重复训练初始化过程. 针对SAR 图像中海上舰船成像小目标的特点改进优化了神经网络结构，实现SAR海面广域舰船目标识别分类算法，并对检测模型进行轻量化压缩处理. 构建SAR图像舰船目标数据集并进行了多次目标检测识别分类实验，体现了提出的检测方法在复杂场景下有可靠的抗干扰能力和准确的目标检测识别性能.

关键词：SAR图像;目标检测;多尺度特征;深度神经网络

中图分类号：TN 957.52                          文献标志码：A

Method of Vessel Target Detection in SAR Images Based

on Multi-scale Feature Superposition

WEI Songjie1，ZHANG Zedong1，XU Zhen1，LIU Meilin2，3，CHEN Wei1

（1. School of Computer Science and Engineering，Nanjing University of Science and Technology，Nanjing 210094，China;

2. School of Aeronautics and Astronautics，Shanghai Jiao Tong University，Shanghai 200240，China;

3. Shanghai Institute of Satellite Engineering，Shanghai 200240，China）

Abstract：For marine vessel target detection in SAR images，deep learning technology alone usually has difficulty in satisfying the detection requirements in accuracy and timeliness. Vessel targets in SAR image are of small sizes and resolutions，which are easily interfered by noise and spot interruption. It is challenging to extract subtle features and eliminate background interference under complex conditions. To overcome the above problems，we propose an end-to-end new detection model based on YOLOv3 framework. The residual module structure is used to avoid network degradation. Combined with deep and shallow feature detection of different target sizes，we extract network parameters for basic features to avoid training from scratch. At the same time，according to the characteristics of small vessel targets in SAR image，the neural network structure is further optimized to achieve fast target detection and categorization in wide-area SAR images，and the detection model is compressed and light-weighted. We construct and utilize a SAR image dataset with different vessel targets for target detection and classification test. The experimental results show that the proposed detection method shows significant anti-jamming ability and detection performance in complex scenes.

Key words：SAR image;target detection;multi-scale features;deep neural network

舰船目标检测在海洋监测和国土安全领域的作用愈发重要，包括海事安全、海上交通、海洋污染防治和边境管制方面具有较为广泛的应用[1]. 合成孔径雷达（Synthetic Aperture Radar，SAR）通过保证相参性的脉冲发射信号，具有较高的分辨率，可以全天时、全天候工作，自问世以来，就在民用遥感、测绘和军事目标探测领域展现出优秀的能力[2]，随着海洋船舶技术的发展，Sentinel-1、ERS系列和高分系列等卫星相继出现，基于SAR图像的舰船目标检……