锚框策略匹配的SSD飞机遥感图像目标检测

doi:10.3778/j.issn.1673-9418.2105108

计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (11): 2596-2608.DOI: 10.3778/j.issn.1673-9418.2105108

锚框策略匹配的SSD飞机遥感图像目标检测

王浩桐¹, 郭中华¹^,²^,⁺()

1.宁夏大学物理与电子电气工程学院，银川 750021
2.宁夏大学沙漠信息智能感知重点实验室，银川 750021

收稿日期:2021-05-27 修回日期:2021-07-19 出版日期:2022-11-01 发布日期:2021-07-26
通讯作者: + E-mail: guozhh@nxu.edu.cn
作者简介:王浩桐（1995—），男，宁夏人，硕士研究生，主要研究方向为计算机视觉、图像处理。
郭中华（1973—），男，山东人，博士，教授，主要研究方向为机器视觉、图像处理。
基金资助:
宁夏自然科学基金(2020AAC03026);宁夏大学研究生创新研究项目(GIP2020075)

Target Detection of SSD Aircraft Remote Sensing Images Based on Anchor Frame Strategy Matching

WANG Haotong¹, GUO Zhonghua¹^,²^,⁺()

1. School of Physics and Electronic and Electrical Engineering, Ningxia University, Yinchuan 750021, China
2. Key Laboratory of Desert Information Intelligent Sensing, Ningxia University, Yinchuan 750021, China

Received:2021-05-27 Revised:2021-07-19 Online:2022-11-01 Published:2021-07-26
About author:WANG Haotong, born in 1995, M.S. candidate. His research interests include computer vision and image processing.
GUO Zhonghua, born in 1973, Ph.D., professor. His research interests include machine vision and image processing.
Supported by:
Natural Science Foundation of Ningxia(2020AAC03026);Graduate Innovative Research Project of Ningxia University(GIP2020075)

摘要/Abstract

摘要：

针对当前飞机遥感图像目标检测算法的精度和实时性不能兼顾的问题，提出了基于SSD的锚框尺度密集化和锚框策略匹配目标检测算法。该算法选用经过改进后的深度残差网络替代SSD算法原有的特征提取网络。结合飞机遥感图像存在小尺度且密集的特点，重新设计了锚框尺度大小、比例和额外增加了一个包含两种尺度的特征层。而后对各个特征层进行锚框密集化操作使得特征层的锚框铺设密度基本相等，提高不同尺度的锚框匹配到真实目标的概率。在不同尺度的正样本锚框数量差距较大的问题上，提出了一种使得不同尺度的正样本锚框数量趋向于总体正样本平均值的锚框策略匹配方法，一定程度上提高训练的有效性和目标检测的鲁棒性。在飞机遥感数据集上进行相关实验，精度均值达到91.15%，每秒帧率为33.4。结果表明，改进后的算法不仅可以在增加较少训练参数的基础上提升检测精度，还能保留SSD算法的实时检测性。

关键词: 目标检测, 遥感图像, 实时检测, 锚框匹配

Abstract:

Aiming at the problem that the accuracy and real-time performance of current aircraft remote sensing image target detection algorithms cannot be balanced, a target detection algorithm based on single shot MultiBox detector (SSD) is proposed for anchor frame scale densification and anchor frame strategy matching. The algorithm uses an improved deep residual network to replace the original feature extraction network of the SSD algorithm. Combined with the small-scale and dense features of aircraft remote sensing images, this paper redesigns the size and proportion of anchor frame and adds a feature layer containing two scales. Then, the anchor frame densification operation is performed on each feature layer to make the anchor frame laying density of the feature layer basically equal, and to improve the probability of matching the anchor frames of different scales to the real target. On the issue of the large gap in the number of positive sample anchor frames of different scales, an anchor frame strategy matching method that makes the number of positive sample anchor frames of different scales tend to the overall positive sample average is proposed, which improves the effectiveness of training and robustness of target detection to a certain extent. Related experiments are conducted on the aircraft remote sensing dataset, the average precision reaches 91.15%, and the frame per second is 33.4. The results show that the improved algorithm can not only increase the detection accuracy on the basis of adding fewer training parameters, but also retain the real-time detec-tability of the SSD algorithm.

Key words: target detection, remote sensing image, real-time detection, anchor box matching

中图分类号:

TP181

王浩桐, 郭中华. 锚框策略匹配的SSD飞机遥感图像目标检测[J]. 计算机科学与探索, 2022, 16(11): 2596-2608.

WANG Haotong, GUO Zhonghua. Target Detection of SSD Aircraft Remote Sensing Images Based on Anchor Frame Strategy Matching[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(11): 2596-2608.

图/表 18

图1 SSD算法网络结构图

Fig.1 Framework diagram of SSD algorithm

图2 残差结构

Fig.2 Residual structure

表1 不同特征提取网络在PASCAL VOC 2007/2012上的检测结果

Table 1 Test results of different feature extraction networks on PASCAL VOC 2007/2012

Training data	Test data	mAP/%
Training data	Test data	VGG16	ResNet101
VOC 2007+VOC 2012	VOC 2007	73.2	76.4
VOC 2007+VOC 2012	VOC 2012	70.4	73.8

图3 ResNet50替换VGG16的网络结构图

Fig.3 Network structure diagram after ResNet50 replaces VGG16

图4 有效感受野变化示意图

Fig.4 Schematic diagram of effective receptive field

表2 锚框的相关信息

Table 2 Information about anchor frame

特征层	特征层尺寸	理论感受野	锚框铺设步长	锚框尺度	锚框宽高比	锚框铺设密度
Conv1	75×75	35×35	4	8,12	1	2,3
Conv2	38×38	187×187	8	24	1, 2/3, 3/5	3
Conv3	19×19	203×203	16	48	1, 2/3, 3/5	3
Conv4	10×10	235×235	30	90	1, 2/3, 3/5	3
Conv5	5×5	299×299	60	180	1, 2/3, 3/5	3
Conv6	3×3	427×427	100	240	1, 2/3, 3/5	3
Conv7	1×1	555×555	300	300	1, 2/3, 3/5	1

图5 锚框密集化示意图

Fig.5 Schematic diagram of anchor frame densification

图6 锚框策略匹配前的结果

Fig.6 Results before anchor box strategy matching

表3 统计信息

Table 3 Statistical information

图6	$L$	$N$	$N$ （不同尺度）	$M$	$D$
（a）	2	12	3,9	6.0	-3,3
（b）	2	33	3,30	16.5	-13.5,13.5

表3 统计信息

Table 3 Statistical information

图6	$L$	$N$	$N$ （不同尺度）	$M$	$D$
（a）	2	12	3,9	6.0	-3,3
（b）	2	33	3,30	16.5	-13.5,13.5

图7 锚框策略匹配后的结果

Fig.7 Results after anchor box strategy matching

图8 匹配前后的平均值

Fig.8 Average value before and after matching

图9 本文数据集的样本示例

Fig.9 Sample examples of dataset in this paper

图10 不同算法的目标检测结果

Fig.10 Target detection results of different algorithms

图11 几种经典目标检测算法的PR曲线

Fig.11 PR curves of several classic target detection algorithms

表4 不同算法在飞机遥感图像上的AP和FPS

Table 4 AP and FPS of different algorithms on aircraft remote sensing images

Method	Backbone	AP/%		FPS
Method	Backbone	IOU=0.50,area=all	IOU=0.50:0.95,area=small	FPS
Faster-RCNN+FPN	VGG16	91.35	40.56	3.4
YOLO	GoogleNet	78.25	28.64	49.6
YOLO v2^[25]	DarkNet19	84.36	33.46	42.8
YOLO v3	DarkNet53	88.14	38.22	38.6
SSD	VGG16	88.72	39.65	41.3
DSSD^[26]	ResNet101	89.55	41.25	26.5
AMDSSD	ResNet50	91.15	41.36	33.4

表5 骨架网络替换的消融实验结果

Table 5 Ablation experiment results of skeleton network replacement

Method	AP/%(IOU=0.50, area=all)
SSD	88.74
SSD+ResNet50(改)	89.36

表6 锚框密集化与锚框策略匹配的消融实验结果

Table 6 Results of ablation experiments matching anchor frame densification and anchor frame strategy

Method	AP/%（IOU=0.50,area=all）
SSD+ResNet50（改）	89.36
SSD+ResNet50（改）+锚框密集	90.58
SSD+ResNet50（改）+锚框密集+锚框策略匹配	91.15

图12 AMDSSD在不同场景下的部分检测结果

Fig.12 Some test results of AMDSSD in different scenarios

参考文献 27

[1]	史文旭, 谭代伦, 鲍胜利. 特征增强SSD算法及其在遥感目标检测中的应用[J]. 光子学报, 2020, 49(1): 148-157.
	SHI W X, TAN D L, BAO S L. Feature enhancement SSD algorithm and its application in remote sensing images target detection[J]. Acta Photonica Sinica, 2020, 49(1): 148-157.
[2]	林娜, 冯丽蓉, 张小青. 基于优化Faster-RCNN的遥感影像飞机检测[J]. 遥感技术与应用, 2021, 36(2): 275-284.
	LIN N, FENG L R, ZHANG X Q. Aircraft detection in remote sensing image based on optimized Faster-RCNN[J]. Remote Sensing Technology and Application, 2021, 36(2): 275-284.
[3]	KRIZHEVSK Y A, SUTSKEVER I, HINTON G. Image-Net classification with deep convolution neural network[C]// Advances in Neural Information Processing Systems 25, Lake Tahoe, Dec 3-6, 2012: 1106-1114.
[4]	GIRSHICK R B. Fast R-CNN[C]// Proceedings of the 2015 IEEE International Conference on Computer Vision, Santiago, Dec 7-13, 2015. Washington: IEEE Computer Society, 2015: 1440-1448.
[5]	REN S, HE K, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intell-igence, 2017, 39(6): 1137-1149.
[6]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]// Procee-dings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, Jun 27-30, 2016. Wash-ington: IEEE Computer Society, 2016: 779-788.
[7]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]// LNCS 9905: Proceedings of the 14th European Conference on Computer Vision, Amsterdam, Oct 11-14, 2016. Cham: Springer, 2016: 21-37.
[8]	董旭彬, 赵清华. 改进Mask R-CNN在航空影像目标检测的研究应用[J]. 计算机工程与应用, 2021, 57(8): 133-144. DOI
	DONG X B, ZHAO Q H. Research and application of improved Mask R-CNN in aerial image target detection[J]. Computer Engineering and Applications, 2021, 57(8): 133-144. DOI
[9]	郭智, 宋萍, 张义, 等. 基于深度卷积神经网络的遥感图像飞机目标检测方法[J]. 电子与信息学报, 2018, 40(11): 2684-2690.
	GUO Z, SONG P, ZHANG Y, et al. Aircraft detection met-hod based on deep convolutional neural network for remote sensing images[J]. Journal of Electronics & Information Technology, 2018, 40(11): 2684-2690.
[10]	王冰, 周焰, 张怀念, 等. 基于改进SSD框架的遥感影像飞机目标检测方法[J]. 火力与指挥控制, 2021, 46(1): 14-19.
	WANG B, ZHOU Y, ZHANG H N, et al. Aircraft detection method based on SSD framework for remote sensing im-ages[J]. Fire Control & Command Control, 2021, 46(1): 14-19.
[11]	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 Recog-nition, Las Vegas, Jun 27-30, 2016. Washington: IEEE Com-puter Society, 2016: 770-778.
[12]	SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. arXiv:1409.1556, 2014.
[13]	黄国新, 李炜, 张比浩, 等. 改进SSD的机场场面多尺度目标检测算法[J]. 计算机工程与应用, 2022, 58(5): 264-270. DOI
	HUANG G X, LI W, ZHANG B H, et al. Improved SSD-based multi-scale object detection algorithm in airport surface[J]. Computer Engineering and Applications, 2022, 58(5): 264-270. DOI
[14]	邹慧海, 侯进. 基于改进SSD算法的道路小目标检测研究[J]. 计算机工程, 2022, 48(5): 281-288.
	ZOU H H, HOU J. Research on road small target detection with improved SSD algorithm[J]. Computer Engineering, 2022, 48(5): 281-288.
[15]	刘建伟, 赵会丹, 罗雄麟, 等. 深度学习批归一化及其相关算法研究进展[J]. 自动化学报, 2020, 46(6): 1090-1120.
	LIU J W, ZHAO H D, LUO X L, et al. Research progress on batch normalization of deep learning and its related algorithms[J]. Acta Automatica Sinica, 2020, 46(6): 1090-1120.
[16]	XU J, LI Z, DU B, et al. Reluplex made more practical: Leaky ReLU[C]// Proceedings of the 2020 IEEE Symposium on Computers and Communications, Rennes, Jul 7-10, 2020. Piscataway: IEEE, 2020: 1-7.
[17]	LUO W J, LI Y J. Understanding the effective receptive field in deep convolutional neural networks[J] arXiv: 1701.04128, 2017.
[18]	陆保国, 梁博, 马焕芳. 光学遥感影像飞机目标识别与分类方法[J]. 指挥信息系统与技术, 2020, 11(5): 78-82.
	LU B G, LIANG B, MA H F. Optical remote sensing image aircraft target recognition and classification method[J]. Command Information System and Technology, 2020, 11(5): 78-82.
[19]	ZHANG S F, WANG X B, LEI Z, et al. FaceBoxes: a CPU real-time and accurate unconstrained face detector[J]. Neuro-computing, 2019, 7: 6-16.
[20]	ZHANG S F, ZHU X Y, LEI Z, et al. S³FD: single shot scale-invariant face detector[J]. arXiv:1708.05237, 2017.
[21]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 99: 2999-3007.
[22]	欧攀, 张正, 路奎, 等. 基于卷积神经网络的遥感图像目标检测[J]. 激光与光电子学进展, 2019, 56(5): 66-72.
	OU P, ZHANG Z, LU K, et al. Object detection of remote sensing images based on convolutional neural networks[J]. Laser & Optoelectronics Progress, 2019, 56(5): 66-72.
[23]	LIN T Y, DOLLAR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recog-nition, Honolulu, Jul 21-26, 2017. Washington: IEEE Com-puter Society, 2017: 936-944.
[24]	REDMON J, FARHADI A. YOLOv3: an incremental improve-ment[J]. arXiv:1804.02767, 2018.
[25]	REDMON J, FARHADI A. YOLO9000: better, faster, stron-ger[C]// Proceedings of the 2017 IEEE Conference on Com-puter Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 6517-6525.
[26]	FU C, LIU W, RANGA A, et al. DSSD: deconvolutional single shot detector[J]. arXiv:1701.06659, 2017.
[27]	李明山, 韩清鹏, 张天宇, 等. 改进SSD的安全帽检测方法[J]. 计算机工程与应用, 2021, 57(8): 192-197. DOI
	LI M S, HAN Q P, ZHANG T Y, et al. Safety helmet detection method of improved SSD[J]. Computer Engine-ering and Applications, 2021, 57(8): 192-197.

锚框策略匹配的SSD飞机遥感图像目标检测

Target Detection of SSD Aircraft Remote Sensing Images Based on Anchor Frame Strategy Matching

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 18

参考文献 27

相关文章 15

编辑推荐

Metrics

[1]	陈灏然, 彭力, 李文涛, 戴菲菲. 加权网络下的小目标检测算法[J]. 计算机科学与探索, 2022, 16(9): 2143-2150.
[2]	任宁, 付岩, 吴艳霞, 梁鹏举, 韩希. 深度学习应用于目标检测中失衡问题研究综述[J]. 计算机科学与探索, 2022, 16(9): 1933-1953.
[3]	彭豪, 李晓明. 多尺度选择金字塔网络的小样本目标检测算法[J]. 计算机科学与探索, 2022, 16(7): 1649-1660.
[4]	孙方伟, 李承阳, 谢永强, 李忠博, 杨才东, 齐锦. 深度学习应用于遮挡目标检测算法综述[J]. 计算机科学与探索, 2022, 16(6): 1243-1259.
[5]	董文轩, 梁宏涛, 刘国柱, 胡强, 于旭. 深度卷积应用于目标检测算法综述[J]. 计算机科学与探索, 2022, 16(5): 1025-1042.
[6]	伏轩仪, 张銮景, 梁文科, 毕方明, 房卫东. 锚点机制在目标检测领域的发展综述[J]. 计算机科学与探索, 2022, 16(4): 791-805.
[7]	赵鹏飞, 谢林柏, 彭力. 融合注意力机制的深层次小目标检测算法[J]. 计算机科学与探索, 2022, 16(4): 927-937.
[8]	王燕妮, 余丽仙. 注意力与多尺度有效融合的SSD目标检测算法[J]. 计算机科学与探索, 2022, 16(2): 438-447.
[9]	阮晨钊, 张祥森, 刘科, 赵增顺. 深度学习的人-物体交互检测研究进展[J]. 计算机科学与探索, 2022, 16(2): 323-336.
[10]	李青援, 邓赵红, 罗晓清, 顾鑫, 王士同. 注意力与跨尺度融合的SSD目标检测算法[J]. 计算机科学与探索, 2022, 16(11): 2575-2586.
[11]	李春标, 谢林柏, 彭力. 特征混合增强与多损失融合的显著性目标检测[J]. 计算机科学与探索, 2022, 16(10): 2395-2404.
[12]	钱伍, 王国中, 李国平. 改进YOLOv5的交通灯实时检测鲁棒算法[J]. 计算机科学与探索, 2022, 16(1): 231-241.
[13]	李科岑, 王晓强, 林浩, 李雷孝, 杨艳艳, 孟闯, 高静. 深度学习中的单阶段小目标检测方法综述[J]. 计算机科学与探索, 2022, 16(1): 41-58.
[14]	李志欣, 陈圣嘉, 周韬, 马慧芳. 协同级联网络和对抗网络的目标检测[J]. 计算机科学与探索, 2022, 16(1): 217-230.
[15]	王迪聪, 白晨帅, 邬开俊. 基于深度学习的视频目标检测综述[J]. 计算机科学与探索, 2021, 15(9): 1563-1577.