基于角度敏感的空间注意力机制的轻量型旋转目标检测器

doi:10.11959/j.issn.2096-6652.202133

智能科学与技术学报 ›› 2021, Vol. 3 ›› Issue (3): 322-333.doi: 10.11959/j.issn.2096-6652.202133

• 专刊：目标智能检测与识别 • 上一篇下一篇

基于角度敏感的空间注意力机制的轻量型旋转目标检测器

尹开石¹, 杨萌², 顾曦³, 王志成³

¹ 东华大学计算机科学与技术学院，上海 201620
² 中国舰船研究设计中心，湖北武汉 430064
³ 同济大学电子与信息工程学院，上海 201804

修回日期:2021-08-06 出版日期:2021-09-15 发布日期:2021-09-01
作者简介:尹开石（2001− ），男，东华大学计算机科学与技术学院在读，主要研究方向为图像分析、机器学习
杨萌（1986− ），男，博士，中国舰船研究设计中心高级工程师，主要研究方向为舰船总体技术、环境智能感知
顾曦（1996− ），女，同济大学电子与信息工程学院硕士生，主要研究方向为机器学习、深度学习
王志成（1975− ），男，博士，同济大学电子与信息工程学院副研究员，主要研究方向为机器学习、深度学习
基金资助:
国防基础科研计划资助项目(JCKY2020206B037)

Light weight rotating object detector based on angle sensitive spatial attention mechanism

Kaishi YIN¹, Meng YANG², Xi GU³, Zhicheng WANG³

¹ School of Computer Science and Technology, Donghua University, Shanghai 201620, China
² China Ship Development and Design Center, Wuhan 430064, China
³ College of Electronic and Information Engineering, Tongji University, Shanghai 201804, China

Revised:2021-08-06 Online:2021-09-15 Published:2021-09-01
Supported by:
The National Defense Basic Research Program of China(JCKY2020206B037)

摘要/Abstract

摘要：

随着深度学习的快速发展，近年来越来越多的基于锚框的目标检测算法被应用于遥感图像上，然而提高算法精度的代价是牺牲了检测速度。因此，选择无锚框的目标检测网络框架，针对遥感场景的特点，提出了一种旋转框的遥感检测算法。根据旋转框与其外接矩形框的空间位置关系，提出了一种简单有效的旋转框表示方式。此外，设计了一种用于辅助检测旋转目标的角度敏感的空间注意力机制，通过引入角度信息提升模型对旋转目标的检测能力。提出的算法在公开遥感数据集DOTA上进行了实验，旋转框的目标检测网络的平均精度均值达到了68.5%，检测速度达到了每秒17.4帧图像。

关键词: 旋转检测, 遥感图像, 无锚框, 注意力机制

Abstract:

With the rapid development of deep learning, more and more target detection algorithms based on anchor frame are applied to remote sensing images in recent years.However, the cost of improving the accuracy of the algorithm is to sacrifice the detection speed.Therefore, the target detection network framework of anchor free was chosen, and a remote sensing detection algorithm of rotating frame was proposed according to the characteristics of remote sensing scene.A simple and effective representation of rotating frame was proposed according to the spatial position relationship between rotating frame and its external rectangular frame.In addition, an angle sensitive attention mechanism was designed to assist the detection of rotating targets.By introducing angle information, the detection ability of the model for rotating targets was improved.The proposed algorithm was tested on the open remote sensing dataset DOTA.The mean average precision of target detection network of rotating frame is 68.5% and the detection speed is 17.4 frames per second.

Key words: rotation detection, remote sensing image, anchor free, attention mechanism

中图分类号:

TP181

尹开石, 杨萌, 顾曦, 等. 基于角度敏感的空间注意力机制的轻量型旋转目标检测器[J]. 智能科学与技术学报, 2021, 3(3): 322-333.

Kaishi YIN, Meng YANG, Xi GU, et al. Light weight rotating object detector based on angle sensitive spatial attention mechanism[J]. Chinese Journal of Intelligent Science and Technology, 2021, 3(3): 322-333.

图/表 16

图1

图2

图3

图4

图5

图6

图7

图8

图9

表1

表2

表3

表4

表5

表6

图10

参考文献 20

[1]	GIRSHICK R , DONAHUE J , DARRELL T ,et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2014: 580-587.
[2]	GIRSHICK R , . Fast R-CNN[C]// Proceedings of the 2015 IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2015: 1440-1448.
[3]	REN S Q , HE K M , GIRSHICK R ,et al. Faster R-CNN:towards real-time object detection with region proposal networks[C]// Proceedings of the IEEE Transactions on Pattern Analysis and Machine Intelligence. Piscataway:IEEE Press, 2015: 1137-1149.
[4]	HE K M , GKIOXARI G , DOLLáR P ,, et al . Mask R-CNN[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2017: 2980-2988.
[5]	HUANG L C , YANG Y , DENG Y F ,et al. DenseBox:unifying landmark localization with end to end object detection[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway:IEEE Press, 2015.
[6]	REDMON J , DIVVALA S , GIRSHICK R ,et al. You only look once:unified,real-time object detection[C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 779-788.
[7]	LAW H , DENG J . CornerNet:detecting objects as paired keypoints[J]. International Journal of Computer Vision, 2020,128(3): 642-656.
[8]	LIN T Y , MAIRE M , BELONGIE S ,et al. Microsoft COCO:common objects in context[M]// Computer Vision-ECCV 2014. Cham: Springer International Publishing, 2014: 740-755.
[9]	ZHOU X Y , WANG D Q , KR?HENBüHL P , . Objects as points[J]. arXiv preprint,2019,arXiv:1904.07850.
[10]	ZHOU X Y , ZHUO J C , KRAHENBUHL P . Bottom-up object detection by grouping extreme and center points[C]// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2019.
[11]	MA T S , TIAN W H , KUANG P ,et al. An anchor-free object detector with novel corner matching method[J]. Knowledge-Based Systems, 2021,224: 107083.
[12]	LI Q M , QIANG H , LI J . Conditional random fields as message passing mechanism in anchor-free network for multi-scale pedestrian detection[J]. Information Sciences, 2021,550: 1-12.
[13]	CHENG G , ZHOU P C , HAN J W . Learning rotation-invariant convolutional neural networks for object detection in VHR optical remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016,54(12): 7405-7415.
[14]	HE K M , ZHANG X Y , REN S Q ,et al. Deep residual learning for image recognition[C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 770-778.
[15]	LIN T Y , DOLLáR P ,, GIRSHICK R , et al . Feature pyramid networks for object detection[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2017: 936-944.
[16]	HU J , SHEN L , SUN G . Squeeze-and-excitation networks[C]// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2018: 7132-7141.
[17]	CUBUK E D , ZOPH B , SHLENS J ,et al. RandAugment:practical automated data augmentation with a reduced search space[C]// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway:IEEE Press, 2020: 3008-3017.
[18]	MA J Q , SHAO W Y , YE H ,et al. Arbitrary-oriented scene text detection via rotation proposals[J]. IEEE Transactions on Multimedia, 2018,20(11): 3111-3122.
[19]	YANG X , LIU Q Q , YAN J C ,et al. R3det:refined single-stage detector with feature refinement for rotating object[J]. arXiv preprint,2019,arXiv:1908.05612.
[20]	JIANG Y Y , ZHU X Y , WANG X B ,et al. R2CNN:rotational region CNN for orientation robust scene text detection[J]. arXiv preprint,2017,arXiv:1706.09579.

类别	方法1	方法2	方法3	方法4
飞机	93.0%	92.2%	71.8%	57.1%
轮船	80.6%	80.1%	14.5%	51.9%
储罐	89.3%	89.7%	83.9%	60.4%
棒球场	67.7%	54.3%	57.4%	37.7%
网球场	94.6%	94.8%	55.7%	51.8%
篮球场	62.3%	50.7%	46.6%	40.6%
地面跑道	53.7%	44.4%	42.3%	41.5%
港口	55.5%	46.5%	19.3%	34.5%
桥梁	29.7%	24.2%	9.0%	23.5%
大型车辆	78.7%	76.5%	9.3%	43.4%
小型车辆	53.4%	57.1%	17.0%	38.8%
直升机	13.8%	25.7%	41.5%	29.6%
环形交叉路口	69.6%	65.3%	64.3%	47.2%
足球场	62.6%	39.4%	47.9%	38.8%
游泳池	52.2%	49.2%	28.1%	35.1%
mAP	63.7%	59.3%	40.5%	42.1%

类别	Resnet50	Restnet50+atten	Resnet101	Resnet101+atten
飞机	88.6%	88.7%	89.0%	88.8%
轮船	79.0%	82.7%	79.2%	84.0%
储罐	86.2%	86.6%	87.1%	85.0%
棒球场	68.6%	72.6%	73.8%	72.8%
网球场	90.7%	90.4%	90.8%	90.7%
篮球场	65.2%	52.2%	63.0%	63.5%
地面跑道	57.7%	55.0%	57.6%	57.1%
港口	63.1%	61.8%	63.0%	61.6%
桥梁	45.5%	40.2%	42.9%	41.9%
大型车辆	80.7%	77.9%	79.4%	80.7%
小型车辆	60.3%	60.5%	61.8%	62.8%
直升机	25.2%	57.7%	54.9%	53.1%
环形交叉路口	66.0%	60.8%	63.7%	66.8%
足球场	66.6%	63.3%	54.3%	56.2%
游泳池	59.8%	59.8%	62.2%	62.8%
mAP	66.8%	67.3%	68.1%	68.5%

算法	特征提取网络	mAP
无通道注意力机制	Resnet101	62.8%
有通道注意力机制	Resnet101	65.1%

算法	特征提取网络	mAP
+多尺度融合	Resnet50	62.4%
无多尺度融合	Resnet50	62.7%
+多尺度融合	Resnet101	64.0%
无多尺度融合	Resnet101	62.8%

算法	mAP	每秒处理帧数/帧	模型大小/MB
Faster R-CNN^[3]	52.9%	12.5	237
R2CNN^[18]	60.6%	0.05	359
RRPN^[20]	61.0%	0.09	369
R3det^[19]	73.7%	1.5	787
本文提出的算法	68.5%	17.4	218

基于角度敏感的空间注意力机制的轻量型旋转目标检测器

Light weight rotating object detector based on angle sensitive spatial attention mechanism

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 16

参考文献 20

相关文章 2

Metrics

推荐阅读 0

旋转框参数表示方法	水平框	旋转框
方法1	64.9%	63.7%
方法2	64.3%	59.3%
方法3	64.4%	40.5%
方法4	64.7%	42.1%

[1]	陈妍, 罗雪琴, 梁伟, 谢永芳. 基于情感信息融合注意力机制的抑郁症识别[J]. 智能科学与技术学报, 2022, 4(4): 600-609.
[2]	沈继锋, 刘岳, 韦浩, 左欣, 杨万扣. 基于差分特征注意力机制的无锚框多光谱行人检测算法[J]. 智能科学与技术学报, 2021, 3(3): 294-303.