基于人眼视点图的特征融合小目标检测算法

doi:10.12305/j.issn.1001-506X.2022.04.07

摘要/Abstract

摘要：

针对目前深度学习小目标检测算法在实际应用中存在的漏检率高、精度低等问题, 提出了一种基于人眼视点图的特征融合小目标检测算法。基于多类别单阶检测器(single shot multibox detection, SSD)算法通过不同扩张率的空洞卷积融合, 在基础网络上获得具有类似人眼感受野的浅层特征层; 对附加网络中的特征层进行信息融合, 合并上下文信息, 增加位置信息和全局语义信息, 从而提升小目标检测精度。通过PASCAL VOC 2007数据集验证, 结果表明, 该算法较传统SSD算法检测精度提升了3.7%, 较改进的小目标检测算法Bi-SSD精度提升了0.8%, 验证了选择更有表征能力的特征层是有效提升小目标检测精度的方法。

关键词: 小目标检测, 单阶检测器算法, 空洞卷积空间金字塔, 特征金字塔融合

Abstract:

Aiming at the high missed detection rate and low accuracy of the current deep learning small target detection algorithm in practical applications, this paper proposes a feature fusion small target detection algorithm based on the human eye view-point map. Based on the single shot multibox detection (SSD) algorithm, through the convolution fusion of holes with different expansion rates, a shallow feature layer similar to the receptive field of the human eye is obtained on the basic network. The feature layer in the additional network performs information fusion, merges context information, adds position information and global semantic information, thereby improving the accuracy of small target detection. Validated by PASCAL VOC 2007 data set, the results show that the detection accuracy of this algorithm is improved by 3.7% compared with the traditional SSD algorithm, and the accuracy of the improved small target detection algorithm Bi-SSD is increased by 0.8%. It is verified that selecting a feature layer with more characterization ability is an effective method to improve the accuracy of small target detection.

Key words: small target detection, single shot multibox detection (SSD) algorithm, hollow convolutional spatial pyramid, feature pyramid fusion

中图分类号:

TP751.1

魏文晓, 刘洁瑜, 沈强, 李成. 基于人眼视点图的特征融合小目标检测算法[J]. 系统工程与电子技术, 2022, 44(4): 1120-1127.

Wenxiao WEI, Jieyu LIU, Qiang SHEN, Cheng LI. Feature fusion small target detection algorithm based on human eye view-point map[J]. Systems Engineering and Electronics, 2022, 44(4): 1120-1127.

图/表 12

图1

图2

图3

图4

图5

图6

表1

图7

表2

表3

表4

图8

参考文献 30

1	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]//Proc. of the European Conference on Computer Vision, 2016: 21-37.
2	SINDHWANI N , VERMA S , BAJAJ T , et al. Comparative analysis of intelligent driving and safety assistance systems using YOLO and SSD model of deep learning[J]. International Journal of Information System Modeling and Design, 2021, 12 (1): 131- 146. doi: 10.4018/IJISMD.2021010107
3	ZHANG S F, WEN L Y, BIAN X, et al. Single-shot refinement neural network for object detection[C]//Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition, 2018: 4203-4212.
4	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 779-788.
5	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 7263-7271.
6	朱明明, 许悦雷, 马时平, 等. 基于特征融合与软判决的遥感图像飞机检测[J]. 光学学报, 2019, 39 (2): 71- 77.
	ZHU M M , XU Y L , MA S P , et al. Airplane detection based on feature fusion and soft decision in remote sensing images[J]. Acta Optica Sinica, 2019, 39 (2): 71- 77.
7	REN S, HE K, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[C]//Proc. of the Advances in Neural Information Processing Systems, 2015: 91-99.
8	朱敏超, 冯涛, 张钰. 基于FD-SSD的遥感图像多目标检测方法[J]. 计算机应用与软件, 2019, 36 (1): 232- 238. doi: 10.3969/j.issn.1000-386x.2019.01.042
	ZHU M C , FENG T , ZHANG Y . Remote sensing image multi-target detection method based on FD-SSD[J]. Computer Applications and Software, 2019, 36 (1): 232- 238. doi: 10.3969/j.issn.1000-386x.2019.01.042
9	LI Z X, ZHOU F Q. FSSD: feature fusion single shot multibox detector[EB/OL]. [2021-01-10]. https://arxiv.org/abs/1712.00960.
10	FU C Y, LIU W, RANGA A, et al. DSSD: deconvolutional single shot detector[EB/OL]. [2021-01-03]. https://arxiv.org/abs/1701.06659.
11	郑浦, 白宏阳, 李伟, 等. 复杂背景下的小目标检测算法[J]. 浙江大学学报(工学版), 2020, 54 (9): 1777- 1784.
	ZHENG P , BAI H Y , LI W , et al. Small target detection algorithm in complex background[J]. Journal of Zhejiang University (Engineering Science), 2020, 54 (9): 1777- 1784.
12	汪能, 胡君红, 刘瑞康, 等. 基于Bi-SSD的小目标检测算法[J]. 计算机系统应用, 2020, 29 (11): 139- 144.
	WANG N , HU J H , LIU R K , et al. Small target detection algorithm based on Bi-SSD[J]. Computer Systems & Applications, 2020, 29 (11): 139- 144.
13	赵彤, 刘洁瑜, 刘星. 空对地场景信息辅助的目标检测算法[J]. 计算机辅助设计与图形学学报, 2019, 31 (10): 1795- 1801.
	ZHAO T , LIU J Y , LIU X . Target detection algorithm using air-to-ground scene information[J]. Journal of Computer-Aided Design and Computer Graphics, 2019, 31 (10): 1795- 1801.
14	WANDELL B A , WINAWER J . Computational neuroimaging and population receptive fields[J]. Trends in Cognitive Sciences, 2015, 19 (6): 349- 357. doi: 10.1016/j.tics.2015.03.009
15	LIU S T, HUANG D. Receptive filed block net for accurate and fast object detection[C]//Proc. of the European Conference on Computer Vision, 2018: 385-400.
16	DAI J F, LI Y, HE K M, et al. R-FCN: object detection via region-based fully convolutional networks[C]//Proc. of the Advances in the Neural Information Processing Systems, 2016: 379-387.
17	ZHOU P, NI B B, GENG C, et al. Scale-transferrable object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 528-537.
18	DANG L F, ZUO W Y, SHI Q. Research on SSD countersink defect detection method based on machine vision[C]//Proc. of the Asta Conference on Geological Research and Environmental Technology, 2021, 632: 042068.
19	IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]//Proc. of the International Conference on Machine Learning, 2015.
20	SZEGEDY C, WEI L, YANG Q J, et al. Going deeper with convolutions[C]//Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition, 2015.
21	SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]//Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition, 2016: 2818-2826.
22	SZEGEDY C, IOFFE S, VANHOUCKE V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]//Proc. of the IEEE International Conference on Computer Vision and Pattern Recognition, 2017.
23	KIM K H, HONG S, ROH B, et al. PVANET: deep but lightweight neural networks for real-time object detection[EB/OL]. [2021-01-03]. https://arxiv.org/abs/1608.08021.
24	DAI J, QI H Z. Deformable convolutional networks[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.
25	HUANG D, WANG Y H. Receptive field block net for accurate and fast object detection[C]//Proc. of the European Conference on Computer Vision, 2018.
26	ZHANG W , SUN W H . Research on small moving target detection algorithm based on complex scene[J]. Journal of Phy-sics: Conference Series, 2021, 1738, 012093. doi: 10.1088/1742-6596/1738/1/012093
27	YANG W, ZHANG J, ZHANG Z B, et al. Research on real-time vehicle detection algorithm based on deep learning[C]//Proc. of the Chinese Conference on Pattern Recognition and Computer Vision, 2018: 126-127.
28	LEE S , CHAN C , MAYO S J , et al. How deep learning extracts and learns leaf features for plant classification[J]. Pattern Recognition, 2017, 71, 118- 131. doi: 10.1016/j.patcog.2017.06.002
29	HE K, GKIOXARI G, DOLLAR P, et al. Mask R-CNN[C]//Proc. of the IEEE International Conference on Computer Vision, 2017: 2961-2969.
30	REDOMN J, FARHADI A. YOLOv3: an incremental improvement[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.

类别	aero	bike	bird	boat	bottle	bus	car	cat	chair	cow
小目标数量	294	424	693	545	1 158	227	2 104	118	2 209	431
总量数	642	807	1 175	791	1 291	526	3 185	759	2 806	685

类别	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv
小目标数量	163	245	230	313	6 258	923	482	146	162	495
总数量	609	1 068	801	759	10 674	1 217	664	821	630	728

方法	mAP	aero	bike	bird	boat	bottle	bus	mbike	person	plant	sheep
Faster-RCNN	69.9	70.0	80.6	70.1	57.3	49.9	78.2	75.0	76.3	39.1	68.3
SSD	68.0	73.4	77.5	64.1	59.0	38.9	75.2	77.0	72.5	41.2	64.2
本文算法	71.7	76.4	79.4	66.5	65.8	44.1	82.4	80.4	74.7	45.0	66.5

方法	car	cat	chair	cow	table	dog	horse	sofa	train	tv	—
Faster-RCNN	80.4	82.0	52.2	75.3	67.2	80.3	79.8	67.3	81.1	67.6	—
SSD	80.8	78.5	46.0	67.8	69.2	76.6	82.1	69.1	78.0	68.5	—
本文算法	83.4	78.9	51.2	78.3	70.4	79.6	83.5	70.2	84.6	72.6	—

空洞卷积空间金字塔模块	特征金字塔融合模块二	特征金字塔融合模块一	mAP/%
—	—	—	77.1
√	—	—	78.3
—	√(无门控)	√	62.7
—	√(门控)	√	77.8
√	√(门控)	√	79.2

方法	mAP	aero	bike	bird	boat	bottle	bus	mbike	person	plant	sheep
Bi-SSD^[7]	78.4	81.5	84.4	75.7	70.7	54.4	86.2	85.5	79.9	52.0	80.0
YOLOv3^[18]	79.2	85.5	85.5	75.6	70.0	66.5	87.6	86.2	83.3	56.2	75.3
本文算法[300]	79.2	81.8	85.2	77.0	72.6	55.0	86.3	85.9	80.2	52.9	79.9
本文算法[512]	81.3	83.6	87.4	82.3	74.0	61.4	88.6	88.3	84.1	57.2	77.4

方法	car	cat	chair	cow	table	dog	horse	sofa	train	tv	—
Bi-SSD^[7]	86.7	88.5	60.9	86.0	77.5	86.8	87.1	79.5	86.6	77.2	—
YOLOv3^[18]	87.7	89.4	64.3	83.5	73.6	85.9	86.9	78.0	86.4	77.8	—
本文算法[300]	86.3	88.1	63.9	85.3	78.3	86.2	87.3	80.5	87.6	78.7	—
本文算法[512]	91.7	88.9	66.3	87.3	79.3	87.6	88.5	77.3	92.0	83.2	—

[1]	时艳玲, 刘子鹏, 张学良, 顾为亮. 基于EMD能量占比的海面漂浮小目标特征检测[J]. 系统工程与电子技术, 2021, 43(2): 300-310.
[2]	赵若辰, 王敬东, 林思玉, 顾东泽. 基于卷积神经网络的小型建筑物检测算法[J]. 系统工程与电子技术, 2021, 43(11): 3098-3106.
[3]	陆福星, 李夜金, 陈忻, 陈桂林, 饶鹏. 基于Top-hat变换的PM模型弱小目标检测[J]. 系统工程与电子技术, 2018, 40(7): 1417-1422.
[4]	林建粦，平西建，马德宝. 基于规范化线性滤波器的漂移扫描恒星抑制[J]. Journal of Systems Engineering and Electronics, 2012, 34(5): 864-870.
[5]	陈炳文，王文伟，秦前清. 基于Fuzzy-ART神经网络的红外弱小目标检测[J]. Journal of Systems Engineering and Electronics, 2012, 34(5): 857-863.
[6]	郑红,隋强强,陈磊. 运动小目标检测最优图像帧数概率模型[J]. Journal of Systems Engineering and Electronics, 2011, 33(1): 8-0012.
[7]	王鑫, 唐振民. 改进的分形算法在弱小目标检测中的应用[J]. Journal of Systems Engineering and Electronics, 2009, 31(9): 2071-2075.
[8]	吴一全, 吴文怡, 罗子娟. 基于最小一乘和遗传算法的红外弱小目标检测[J]. Journal of Systems Engineering and Electronics, 2009, 31(3): 575-578.
[9]	陈尚锋, 肖山竹, 卢焕章. 图像序列弱小目标能量积累检测研究[J]. Journal of Systems Engineering and Electronics, 2009, 31(2): 288-291.