Remote Sensing for Natural Resources >
Detecting ships from SAR images based on high-dimensional contextual attention and dual receptive field enhancement
Received date: 2024-01-26
Revised date: 2024-11-21
Online published: 2026-06-03
The abundant contextual information in synthetic aperture radar (SAR) images remains underutilized in deep learning-based ship detection. Hence, this study proposed a novel method for detecting ships from SAR images based on high-dimensional contextual attention and dual receptive field enhancement. The dual receptive field enhancement was employed to extract multi-dimensional feature information from SAR images, thereby guiding the dynamic attention matrix to learn rich contextual information during the coarse-to-fine extraction of high-dimensional features. Based on YOLOv7, a YOLO-HD network was constructed by incorporating a lightweight convolutional module, a lightweight asymmetric multi-level compression detection head, and a new loss function,XIoU. A comparative experiment was conducted on the E-HRSID and SSDD datasets. The proposed method achieved average detection accuracy of 91.36 % and 97.64 %, respectively, representing improvements by 4.56 and 9.83 percentage points compared to the original model, and outperforming other classical models.
GUO Wei , LI Yu , JIN Haibo . Detecting ships from SAR images based on high-dimensional contextual attention and dual receptive field enhancement[J]. Remote Sensing for Natural Resources, 2025 , 37(3) : 104 -112 . DOI: 10.6046/zrzyyg.2024047
表1 E-HRSID和SSDD数据集对比实验结果Tab.1 Comparison experiment results of E-HRSID and SSDD datasets |
| 模型 | E-HRSID | SSDD | ||||||
|---|---|---|---|---|---|---|---|---|
| P/% | R/% | mAP/% | F1 | P/% | R/% | mAP/% | F1 | |
| CenterNet | 96.63 | 60.95 | 76.77 | 0.75 | 97.46 | 75.42 | 89.04 | 0.83 |
| Efficientdet | 97.29 | 22.36 | 34.73 | 0.36 | 95.77 | 25.09 | 73.83 | 0.40 |
| Faster R-CNN | 34.3 | 35.71 | 26.95 | 0.35 | 73.50 | 68.07 | 88.12 | 0.71 |
| RetinaNet | 93.31 | 27.65 | 34.35 | 0.43 | 86.81 | 63.14 | 80.86 | 0.73 |
| SSD | 88.79 | 16.34 | 40.64 | 0.28 | 95.80 | 42.07 | 89.58 | 0.58 |
| YOLOv7 | 88.16 | 78.16 | 86.80 | 0.83 | 90.80 | 78.00 | 87.81 | 0.84 |
| YOLOv8 | 89.53 | 83.27 | 90.47 | 0.86 | 95.40 | 91.80 | 97.10 | 0.94 |
| YOLO-HD | 90.65 | 84.36 | 91.36 | 0.87 | 95.25 | 95.55 | 97.64 | 0.95 |
表3 消融试验Tab.3 Ablation experiment |
| 模型 | DRFE | HD-ELAN | LAMCD | XIoU | L-ELAN | P/% | R/% | mAP/% | 参数量/MB | GFLOPS |
|---|---|---|---|---|---|---|---|---|---|---|
| 基础模型 | — | — | — | — | — | 88.16 | 78.16 | 86.80 | 38.4 | 105.4 |
| Net1 | √ | — | — | — | — | 88.14 | 81.39 | 88.85 | 43.4 | 108.7 |
| Net2 | — | √ | — | — | — | 90.54 | 81.10 | 89.79 | 39.4 | 162.1 |
| Net3 | √ | √ | — | — | — | 90.45 | 78.90 | 89.29 | 37.6 | 141.2 |
| Net4 | √ | √ | √ | — | — | 91.21 | 82.92 | 90.89 | 51.0 | 187.0 |
| 本文模型 | √ | √ | √ | √ | √ | 90.65 | 84.36 | 91.36 | 56.8 | 143.2 |
| [1] |
曾涛, 温育涵, 王岩, 等. 合成孔径雷达参数化成像技术进展[J]. 雷达学报, 2021, 10(3):327-341.
|
| [2] |
|
| [3] |
|
| [4] |
张帆, 陆圣涛, 项德良, 等. 一种改进的高分辨率SAR图像超像素CFAR舰船检测算法[J]. 雷达学报, 2023, 12(1):120-139.
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
张鹏, 张嘉峰, 刘涛. 基于相干度优化的极化顺轨干涉SAR慢小目标CFAR检测[J]. 北京航空航天大学学报, 2019, 45(3):575-587.
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
/
| 〈 |
|
〉 |