基于关键点的遥感图像舰船目标检测

doi:10.12305/j.issn.1001-506X.2022.08.07

摘要/Abstract

摘要：

针对当前舰船目标检测算法存在锚框遍历计算成本高和特征旋转适应性不足等问题, 提出基于关键点的遥感图像舰船目标检测方法, 通过预估舰船中心点实现目标检测。首先, 引入深度可分离卷积降低参数冗余, 结合SimAM无参注意力机制, 增强对舰船目标的关注度。其次, 引入方向不变模型(orientation-invariant model, OIM)生成方向不变特征图, 增强网络对旋转目标的适应能力。最后, 考虑到遥感图像舰船目标任意方向密集排列, 但舰船目标中心点不变的特点, 采用直接预测目标的中心点, 再回归偏移量、目标尺度和角度的思路, 摆脱锚框遍历机制, 提高检测速度。在HRSC2016和RFUE2021数据集上进行对比实验, 实验结果充分说明了本文方法的有效性和先进性。

关键词: 任意方向舰船检测, 中心点估计, SimAM注意力, 方向不变模型

Abstract:

Problems with current ship target detection method, such as high computational cost of anchor frame traversal; and the rotation invariance of the features extracted from the backbone network is weak and cannot adapt to the ship targets in any direction, resulting in inconsistency. Therefore, a ship target detection method based on key points in remote sensing images is proposed, and the target detection is realized by predicting the ship center point. First, the depth separable convolution is added to reduce the parameter redundancy, and the attention to the ship target is enhanced combined with SimAM nonparametric attention. Second, the orientation-invariant model (OIM) is introduced to generate the orientation-invariant feature map to enhance the adaptability of the network to target rotation. Finally, considering that the ship targets in remote sensing images are densely arranged in any direction, but the center point of the ship target is constant, the idea of directly predicting the center point of the target, and then regressing the offset, target scale and angle is adopted to get rid of the anchor frame traversal mechanism and improve the detection speed. A comparative experiment was conducted on the HRSC2016 and RFUE2021 datasets, and the experimental results fully demonstrate the effectiveness of the proposed method.

Key words: arbitrary-oriented ship detection, center point estimation, SimAM attention, orientation-invariant model (OIM)

中图分类号:

张涛, 杨小冈, 卢瑞涛, 谢学立, 刘闯. 基于关键点的遥感图像舰船目标检测[J]. 系统工程与电子技术, 2022, 44(8): 2437-2447.

Tao ZHANG, Xiaogang YANG, Ruitao LU, Xueli XIE, Chuang LIU. Key-point based method for ship detection in remote sensing images[J]. Systems Engineering and Electronics, 2022, 44(8): 2437-2447.

图/表 17

图1

图2

图3

图4

图5

图6

图7

表1

表2

图8

图9

表3

图10

表4

表5

图11

表6

参考文献 31

1	DENG Z P , SUN H , ZHOU S L , et al. Learning deep ship detector in SAR images from scratch[J]. IEEE Trans.on Geo-science and Remote Sensing, 2019, 1- 19.
2	张玉莲. 光学图像海面舰船目标智能检测与识别方法研究[D]. 长春: 中国科学院大学, 2021.
	ZHANG Y L. Research on intelligent detection and recognition methods of ship targets on the sea surface in optical images[D]. Changchun: Chinese Academy of Science, 2021.
3	LIU S C , DU Q , TONG X H , et al. Unsupervised change detection in multispectral remote sensing images via spectral spatial band expansion[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019, 3578- 3587.
4	张筱晗, 姚力波, 吕亚飞, 等. 基于中心点的遥感图像多方向舰船目标检测[J]. 光子学报, 2020, 49 (4): 210- 218.
	ZHANG X H , YAO L B , LYU Y F , et al. Center based model for arbitrary-oriented ship detection in remote sensing images[J]. Acta Photonica Sinica, 2020, 49 (4): 210- 218.
5	史文旭, 江金洪, 鲍胜利. 基于特征融合的遥感图像舰船目标检测方法[J]. 光子学报, 2020, 49 (7): 57- 67.
	SHI W X , JIANG J H , BAO S L . Ship detection method in remote sensing image based on feature fusion[J]. Acta Photonica Sinica, 2020, 49 (7): 57- 67.
6	张筱晗, 姚力波, 吕亚飞, 等. 双向特征融合的数据自适应SAR图像舰船目标检测模型[J]. 中国图象图形学报, 2020, 25 (9): 1943- 1952.
	ZHANG X H , YAO L B , LYU Y F , et al. Data-adaptive single-shot ship detector with a bidirectional feature fusion module for SAR images[J]. Journal of Image and Graphics, 2020, 25 (9): 1943- 1952.
7	赵江洪, 张晓光, 杨璐, 等. 深度学习的遥感影像舰船目标检测[J]. 测绘科学, 2020, 45 (3): 110- 116.110-116, 134
	ZHAO J H , ZHANG X G , YANG L , et al. Deep learning based detection method for remote sensing images[J]. Science of Surveying and Mapping, 2020, 45 (3): 110- 116.110-116, 134
8	陈冬, 句彦伟. 基于改进型YOLOv3的SAR图像舰船目标检测[J]. 系统工程与电子技术, 2021, 43 (4): 937- 943.
	CHENG D , JU Y W . Ship detection in SAR image based on improved YOLOv3[J]. Systems Engineering and Electronics, 2021, 43 (4): 937- 943.
9	仲伟峰, 郭峰, 向世明, 等. 旋转矩形区域的遥感图像舰船目标检测模型[J]. 计算机辅助设计与图形学学报, 2019, 31 (11): 1935- 1945.
	ZHONG W F , GUO F , XIANG S M , et al. Ship detection in remote sensing based with rotated rectangular region[J]. Journal of Computer-Aided Design & Computer Graphics, 2019, 31 (11): 1935- 1945.
10	LIU Z K, HU J G, WENG L B, et al. Rotated Region-based CNN for ship detection[C]//Proc. of the IEEE International Conference on Image Processing, 2017: 900-904.
11	LIU L, PAN Z, LEI B. Learning a rotation invariant detector with rotatable bounding box[EB/OL]. https://arxiv.org/abs/1711.09405v1.
12	DING J, XUE N, LONG Y, et al. Learning roI transformer for oriented object detection in aerial images[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 2844-2853.
13	ZHANG Z H , GUO W W , ZHU S N , et al. Toward arbitrary-oriented ship detection with rotated region proposal and discrimination networks[J]. IEEE Geoscience & Remote Sensing Letters, 2018,
14	WEI H R , ZHANG Y , CHANG Z H , et al. Oriented objects as pairs of middle lines[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 169, 268- 279. doi: 10.1016/j.isprsjprs.2020.09.022
15	WEI H R, ZHANG Y, WANG B, et al. X-linenet: detecting aircraft in remote sensing images by a pair of intersecting line segments[EB/OL]. [2021-06-30]. https://arxiv.org/abs/1907.12474.
16	ZHOU X Y, WANG D Q, KRHENBVHL P. CenterNet: objects as points[EB/OL]. [2021-09-01]. https://arxiv.org/abs/1904.07850v1.
17	ZHOU Y Z, YE Q X, QIU Q, et al. Oriented response networks[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 4961-4970.
18	YANG L X, ZHANG R Y, LI L D, et al. SimAM: a simple, parameter-free attention module for convolutional neural networks[C]//Proc. of the 38th International Conference on Machine Learning, 2021: 11863-11874.
19	YU F, WANG D Q, SHELHAMER E, et al. Deep layer aggregation[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018: 2403-2413.
20	NEWELL A, YANG K, JIA D. Stacked hourglass networks for human pose estimation[C]//Proc. of the European Conference on Computer Vision, 2016.
21	HU J , SHEN L , ALBANIE S , et al. Squeeze-and-excitation networks[J]. IEEE Trans.on Pattern Analysis and Machine Intelligence, 2017, 42 (8): 2011- 2023.
22	WOO S , PARK J , LEE J Y , et al. CBAM: convolutional block attention module[J]. European Conference on Computer Vision, 2018, 3- 19.
23	ZHANG F, WANG X Y, ZHOU S L, et al. Arbitrary-oriented ship detection through center-head point extraction[EB/OL]. [2021-06-30]. https://arxiv.org/abs/2101.11189.
24	LI M J, GUO W W, ZAHNG Z H, et al. Rotated region based fully convolutional network for ship detection[C]//Proc. of the IEEE International Geoscience and Remote Sensing Symposium, 2018: 673-676.
25	LIN T, GOYAL P, GIRSHICK R B, et al. Focal loss for dense object detection[C]//Proc. of the IEEE International Conference on Computer Vision, 2017: 2999-3007.
26	YANG X, YANG J, YAN J, et al. Scrdet: towards more robust detection for small, cluttered and rotated objects[C]//Proc. of the IEEE/CVF International Conference on Computer Vision, 2019: 8231-8240.
27	YANG X, YAN J. Arbitrary-oriented object detection with circular smooth label[EB/OL]. [2021-09-01]. https://arxiv.org/abs/2003.05597,2020.
28	YANG X, YAN J C, FENG Z M, et al. R3det: refined single-stage detector with feature refinement for rotating object[C]//Proc. of the 35th AAAI Conference on Artificial Intelligence, 2019: 3163-3171.
29	QIAN W, YANG X, PENG S, et al. Learning modulated loss for rotated object detection[C]//Proc. of the 35th AAAI Conference on Artificial Intelligence, 2021: 2458-2466
30	HAN J M , DING J , LI J , et al. Align deep features for oriented object detection[J]. IEEE Trans.on Geoscience and Remote Sensing, 2021,
31	MA J Q , SHAO W Y , YE H , et al. Arbitrary-oriented scene text detection via rotation proposals[J]. IEEE Trans.on Multimedia, 2018,

实验参数	参数配置
优化器	Adam
初始学习率	5×4e
Epoch	230
初始降采样率	R=4

算法	输入大小	mAP_0.5	mAP_0.7	F1	FPS
Res18_dcn	512×512	62.34	58.70	65.12	34.2
DLASeg	512×512	71.52	65.34	74.98	44.3
Hourglass	512×512	73.18	68.46	78.67	13.6
Ours	512×512	77.96	72.09	81.56	41.2
Ours	1 024×1 024	80.57	75.42	83.37	15.4

方法	主干网络	输入大小/像素	mAP_0.5	mAP_0.7	FPS
R²CNN^[24]	Resnet50	512×512	75.09	63.83	10.3
RetinaNet-Rbb^[25]	Resnet50	512×512	70.49	62.82	35.6
SCRDet^[26]	Resnet50	512×512	72.90	63.04	9.2
CSL^[27]	Resnet50	512×512	70.73	61.25	10.4
R³Det^[28]	Resnet50	512×512	67.47	60.17	14.0
RSDet^[29]	Resnet50	512×512	70.74	61.52	15.4
S²A-Net^[30]	Resnet50	512×512	77.19	72.65	33.1
CenterNet-Rbb	DLA34	512×512	71.82	67.52	41.09
CenterNet-Rbb	Hourglass	512×512	72.13	69.21	13.7
Ours¹⁾	DLASeg	512×512	77.96	72.09	41.23

方法	主干网络	输入尺寸/像素	mAP
R²CNN^[24]	Resnet101	800×800	73.1
PPRN^[31]	Resnet101	800×800	79.1
R²PN^[13]	VGG16	800×800	79.6
ROI-Trans^[12]	Resnet101	512×800	86.2
RSDet^[29]	Resnet101	800×800	86.5
CenterNet-Rbb	Hourglass	1 024×1 024	80.1
Ours¹⁾	DLASeg	1 024×1 024	86.7

Baseline	DSC	SinAM	OIM	mAP_0.5	FPS
√				71.82	41.09
√	√			70.09	43.56
√	√	√		73.38	42.89
√	√	√	√	77.96	41.23