欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2024, Vol. 41 ›› Issue (6): 776-785.DOI: 10.7523/j.ucas.2023.017

• 电子信息与计算机科学 • 上一篇    

一种面向SAR图像快速舰船检测的轻量化网络

周文雪, 张华春   

  1. 中国科学院空天信息创新研究院, 北京 100190;中国科学院大学电子电气与通信工程学院, 北京 100049
  • 收稿日期:2022-11-16 修回日期:2023-03-01 发布日期:2023-03-21
  • 通讯作者: 张华春,E-mail:hczhang@mail.ie.ac.cn
  • 基金资助:
    国家自然科学基金(61901445)和北京市自然科学基金(4192065)资助

Lightweight network for fast ship detection in SAR images

ZHOU Wenxue, ZHANG Huachun   

  1. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China;School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2022-11-16 Revised:2023-03-01 Published:2023-03-21

摘要: 在基于深度学习的合成孔径雷达(SAR)图像舰船检测领域,传统的模型通常结构复杂、计算量大,难以适配低算力平台并实现实时检测;同时,依赖于预设锚框的卷积神经网络因锚框位置较难合理设置,容易导致大量计算冗余。针对上述问题,提出一种基于无锚框的端到端轻量化卷积神经网络,设计了一种轻量的通道注意力模块(EESE)并将其应用于解耦合检测头(ED-head)上,有效解决了分类和定位2种任务的冲突。此外,提出一种优化的EIOU损失函数,在保证推理速度几乎不变的情况下有效提升网络性能。在SSDD数据集上的实验结果表明:与YOLOX-nano相比,该方法的AP50和AP分别提高2.1和7.4个百分点,在CPU上推理延迟仅5.33 ms,远小于YOLOX-nano的13.13 ms,实现了精度与效率的平衡。

关键词: 合成孔径雷达(SAR), 舰船检测, 深度学习, 轻量化网络, 无锚框目标检测

Abstract: In the field of SAR image ship detection based on deep learning, traditional models are usually complex in structure and require a large amount of calculation, making them unsuitable for low computing power platforms and real-time detection. And convolutional neural networks that rely on preset anchor boxes will lead to a lot of computational redundancy due to the difficulty of setting a reasonable anchor box. To solve these problems, an end-to-end lightweight convolutional neural network based on anchor-free design is proposed, and a lightweight channel attention module (EESE) is designed and applied to the detection head (ED-head), to resolve the conflict between classification and localization tasks. In addition, an optimized EIOU loss function is proposed, which enables the model to effectively improve the network performance without increasing the inference time. The proposed method is tested on the SSDD dataset, and the experimental results show that compared to YOLOX-nano, AP50 and AP are increased by 2.1 and 7.4 percentage points, respectively, with the CPU latency being only 5.33 ms, much less than 13.13 ms of YOLOX-nano. The proposed method achieves a balance between accuracy and efficiency.

Key words: synthetic aperture radar (SAR), ship detection, deep learning, lightweight network, anchor-free target detection

中图分类号: