采用神经网络架构搜索的高分辨率遥感影像目标检测

doi:10.13229/j.cnki.jdxbgxb.20221472

摘要/Abstract

摘要：

针对传统遥感影像目标检测的深度学习网络需要人工设计、过度依赖专家经验、费力耗时等问题，提出了一种基于神经网络架构搜索的遥感影像目标检测方法，通过逐路径采样和进化搜索策略自动构建高效的目标检测网络，完成遥感影像目标检测任务。在DIOR数据集和RSOD数据集上进行了实验，目标检测平均精度达到67.8%和85.5%，FLOPs为208.47 G和201.67 G，在检测精度和计算效率方面均优于Faster R-CNN、RetinaNet、NAS-FCOS、ResNet Strikes Back、HRNet和GRoIE等现有网络模型。实验结果表明，本方法能自动搜索出高分辨率遥感影像目标检测的网络架构，具有比人工设计的经典网络更优越的性能。

关键词: 遥感, 高分辨率遥感影像, 神经网络架构搜索, 目标检测, 逐路径采样

Abstract:

Aiming at the problems of traditional object detection methods on remote sensing images based on deep learning networks need hand-crafted architectures， which are overly dependent on expert experience and time-consuming， an object detection method for remote sensing images based on neural architecture search is proposed. The network is automatically built by pathwise sampling and evolutionary search strategy for object detection of remote sensing images. Experiments on DIOR dataset and RSOD dataset show that the mean average precision of object detection reached 67.8% and 85.5%， and the FLOPs are 208.47 G and 201.67 G， which are better than the network models such as Faster R-CNN， RetinaNet， NAS-FCOS， ResNet Strikes Back， HRNet and GRoIE in terms of detection accuracy and computational efficiency. The proposed method can automatically search the network architecture for object detection of high-resolution remote sensing images， which is superior to the hand-crafted classical networks.

Key words: remote sensing, high-resolution remote sensing image, network architecture search, object detection, pathwise sampling

中图分类号:

TP751.1

杨军,韩鹏飞. 采用神经网络架构搜索的高分辨率遥感影像目标检测[J]. 吉林大学学报(工学版), 2024, 54(9): 2646-2657.

Jun YANG,Peng-fei HAN. Object detection of high-resolution remote sensing images by neural architecture search[J]. Journal of Jilin University(Engineering and Technology Edition), 2024, 54(9): 2646-2657.

图/表 13

图1

图2

图3

表1

表2

图4

表3

图5

图6

表4

表5

图7

表6

参考文献 26

1	周鹏, 杨军. 采用神经网络架构搜索的遥感影像分割方法[J]. 西安电子科技大学学报, 2021, 48(5): 47-57, 77.
	Zhou Peng, Yang Jun. Semantic segmentation of remote sensing images based on neural architecture search[J]. Journal of Xidian University, 2021, 48(5): 47-57, 77.
2	张晓东, 张力飞, 陈关州, 等. 基于深度学习的遥感影像地物目标检测和轮廓提取一体化模型[J]. 测绘地理信息, 2019, 44(6): 1-5.
	Zhang Xiao-dong, Zhang Li-fei, Chen Guan-zhou, et al. An integrated model of object detection and contour extraction based on deep learning [J]. Journal of Geomatics, 2019, 44(6): 1-5.
3	田婷婷, 杨军. 基于多尺度特征融合网络的遥感影像目标检测[J]. 激光与光电子学进展, 2022, 59(16): 427-435.
	Tian Ting-ting, Yang Jun. Object detection for remote sensing image using multi-scale feature fusion network[J]. Laser & Optoelectronics Progress, 2022, 59(16): 427-435.
4	Cao Y, Niu X, Dou Y. Region-based convolutional neural networks for object detection in very high-resolution remote sensing images[C]∥IEEE International Conference on Natural Computation, Hawaii, USA, 2016: 548-554.
5	Li K, Cheng G, Bu S, et al. Rotation-insensitive and context-augmented object detection in remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(4): 2337-2348.
6	Zhong Y, Han X, Zhang L. Multi-class geospatial object detection based on a position-sensitive balancing framework for high spatial resolution remote sensing imagery[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2018, 138: 281-294.
7	Chen Z, Zhang T, Ouyang C. End-to-end airplane detection using transfer learning in remote sensing images[J]. Remote Sens, 2018, 10: No.139.
8	Liu W, Ma L, Chen H. Arbitrary-oriented ship detection framework in optical remote-sensing images [J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(6): 937-41.
9	Zoph B, Le Q V. Neural architecture search with reinforcement learning[DB/OL]. [2016-11-05].
10	Zoph B, Vasudevan V, Shlens J, et al. Learning transferable architectures for scalable image recognition[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 8697-8710.
11	Real E, Aggarwal A, Huang Y, et al. Regularized evolution for image classifier architecture search[C]∥Proceedings of the AAAI Conference on Artificial Intelligence, Hawaii, USA, 2019: 4780-4789.
12	Liu H, Simonyan K, Yang Y. DARTS: differentiable architecture search[C]∥International Conference on Learning Representations, Vancouver, Canada, 2018: 6-9.
13	Ghiasi G, Lin T Y, Le Q V. NAS-FPN: learning scalable feature pyramid architecture for object detection [C] ∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 7036-7045.
14	Wang N, Gao Y, Chen H, et al. NAS-FCOS: fast neural architecture search for object detection [C] ∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 11940-11948.
15	Xu A, Yao A, Li A, et al. Auto-FPN: automatic network architecture adaptation for object detection beyond classification[C]∥IEEE/CVF International Conference on Computer Vision, Venice, Italy, 2020: 6648-6657.
16	Cao L, Zhang X, Wang Z. Arbitrary-oriented object detection on high-resolution images based on differentiable architecture search [J]. Canadian Journal of Remote Sensing, 2021, 47(5): 719-30.
17	Ma N N, Zhang X Y, Zheng H T, et al. ShuffleNet V2: practical guidelines for efficient CNN architecture design[J/OL]. [2022-11-12].
18	Xia G S, Bai X, Ding J, et al. DOTA: a large-scale dataset for object detection in aerial images[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 3974-3983.
19	Li K, Wan G, Cheng G, et al. Object detection in optical remote sensing images: a survey and a new benchmark[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 159: 296-307.
20	Long Y, Gong Y, Xiao Z, et al. Accurate object localization in remote sensing images based on convolutional neural networks [J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55(5): 2486-2498.
21	Xiao Z, Liu Q, Tang G, et al. Elliptic Fourier transformation-based histograms of oriented gradients for rotationally invariant object detection in remote-sensing images[J]. International Journal of Remote Sensing, 2015, 36(2): 618-644.
22	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 Intelligence, 2017, 39(6): 1137-1149.
23	Rossi L, Karimi A, Prati A. A novel region of interest extraction layer for instance segmentation[C]∥IEEE International Conference on Pattern Recognition, Chengdu, China, 2021: 2203-2209.
24	Sun K, Xiao B, Liu D, et al. Deep high-resolution representation learning for human pose estimation[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 5693-5703.
25	Wightman R, Touvron H, Jégou H. Resnet strikes back: an improved training procedure in timm [DB/OL].[2021-10-01].
26	Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]∥IEEE International Conference on Computer Vision, Venice, Italy, 2017: 2980-2988.

相关文章 15

[1]	朱圣杰,王宣,徐芳,彭佳琦,王远超. 机载广域遥感图像的尺度归一化目标检测方法[J]. 吉林大学学报(工学版), 2024, 54(8): 2329-2337.
[2]	游新冬,郭磊,韩晶,吕学强. 一种工件表面压印字符识别网络[J]. 吉林大学学报(工学版), 2024, 54(7): 2072-2079.
[3]	高云龙,任明,吴川,高文. 基于注意力机制改进的无锚框舰船检测模型[J]. 吉林大学学报(工学版), 2024, 54(5): 1407-1416.
[4]	陈仁祥,胡超超,胡小林,杨黎霞,张军,何家乐. 基于改进YOLOv5的驾驶员分心驾驶检测[J]. 吉林大学学报(工学版), 2024, 54(4): 959-968.
[5]	张云佐,郭威,李文博. 遥感图像密集小目标全方位精准检测算法[J]. 吉林大学学报(工学版), 2024, 54(4): 1105-1113.
[6]	王宏志,宋明轩,程超,解东旋. 基于改进YOLOv4-tiny算法的车距预警方法[J]. 吉林大学学报(工学版), 2024, 54(3): 741-748.
[7]	李雄飞,宋紫萱,朱芮,张小利. 基于多尺度融合的遥感图像变化检测模型[J]. 吉林大学学报(工学版), 2024, 54(2): 516-523.
[8]	王春华,李恩泽,肖敏. 多特征融合和孪生注意力网络的高分辨率遥感图像目标检测[J]. 吉林大学学报(工学版), 2024, 54(1): 240-250.
[9]	蔡志丹,方明,李喆,许佳路. 基于高斯曲率和加权图总变分正则化的遥感图像盲去模糊算法[J]. 吉林大学学报(工学版), 2023, 53(9): 2649-2658.
[10]	朱俊清,赵学儒,马涛,黄晓明,朱洪洲. 基于卫星遥感的路域地质灾害监测方法[J]. 吉林大学学报(工学版), 2023, 53(6): 1861-1872.
[11]	薛珊,张亚亮,吕琼莹,曹国华. 复杂背景下的反无人机系统目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 891-901.
[12]	陶博,颜伏伍,尹智帅,武冬梅. 基于高精度地图增强的三维目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 802-809.
[13]	成丽波,李新月,李喆,贾小宁. 基于曲波变换与拟合优度检验的遥感图像去噪方法[J]. 吉林大学学报(工学版), 2023, 53(11): 3207-3213.
[14]	马为駽,张䶮,马传香,朱飒. 不同光照条件下含噪遥感图像边缘检测算法[J]. 吉林大学学报(工学版), 2023, 53(1): 241-247.
[15]	朱冰,李紫薇,李奇. 基于改进SegNet的遥感图像建筑物分割方法[J]. 吉林大学学报(工学版), 2023, 53(1): 248-254.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

网络	mAP	AP₅₀	AP₇₅	AP_s	AP_m	AP_l
Faster R-CNN	66.1	88.8	74.2	12.2	50.8	76.2
GRoIE	64.1	91.2	72.7	26.0	57.3	71.4
HRNet	63.3	91.1	72.7	28.5	52.0	70.8
ResNet-SB	52.0	76.4	57.5	9.7	31.2	62.1
RetinaNet	34.3	55.1	36.7	3.9	19.4	42.8
NAS-FCOS	55.7	80.4	60.6	10.4	35.8	65.5
本文	67.8	91.3	78.3	36.1	60.6	74.4

类别	Faster R-CNN	GRoIE	HRNet	ResNet-SB	RetinaNet	NAS-FCOS	本文
机场	66.0	60.0	63.4	42.2	10.9	52.2	58.2
车辆	46.7	52.3	51.9	42.0	30.0	47.0	53.0
篮球场	81.0	77.9	72.3	63.9	50.1	68.7	81.9
田径场	82.2	78.1	80.2	65.6	46.9	65.5	81.4
风车	56.1	54.0	47.5	43.2	30.5	43.3	56.5
船舶	47.0	48.5	52.0	43.1	31.6	48.6	48.7
高速公路收费站	74.3	72.0	72.0	63.3	50.5	57.4	76.4
网球场	90.4	86.2	78.4	79.9	75.4	82.0	89.5
高尔夫球场	61.8	53.6	61.0	43.9	32.2	56.3	62.2
立交桥	57.0	59.0	57.1	44.3	22.2	38.9	61.9
储油罐	67.3	70.6	67.1	62.5	54.1	71.3	71.0
棒球场	86.5	85.2	84.7	77.6	75.9	83.5	85.4
港口	41.7	34.3	44.3	26.3	8.4	34.0	51.4
体育场	77.9	76.6	75.9	69.2	47.0	67.9	81.2
桥梁	43.7	49.4	45.4	31.8	10.5	27.5	55.3
飞机	81.5	78.0	74.9	67.8	59.9	75.5	80.6
火车站	49.2	48.9	44.2	19.2	5.2	25.4	52.1
高速公路服务区	70.1	62.8	63.7	44.0	9.6	53.1	68.7
水坝	55.0	52.8	48.4	36.5	19.3	35.4	58.5
烟囱	81.7	81.3	81.6	74.7	63.2	80.0	82.0

网络	FLOPs/G	参数/M	搜索时间/h	训练时间/h	mAP
Faster R-CNN	798.04	32.95	—	20.6	66.1
GRoIE	545.00	43.12	—	22.1	64.1
HRNet	298.65	46.97	—	21.1	63.3
ResNet-SB	216.40	41.22	—	19.3	52.0
RetinaNet	285.98	20.01	—	17.1	34.3
NAS-FCOS	230.62	38.15	11.4	18.4	55.7
本文	208.47	25.38	12.6	20.8	67.8

网络	mAP	AP₅₀	AP₇₅	AP_s	AP_m	AP_l
Faster R-CNN	84.9	99.5	96.4	61.0	83.0	88.0
GRoIE	84.5	99.4	96.3	60.7	82.2	87.9
HRNet	84.1	99.2	97.3	59.9	82.5	87.2
ResNet-SB	77.4	97.5	90.5	23.3	66.4	82.3
RetinaNet	51.4	84.8	55.8	33.6	52.5	54.0
NAS-FCOS	77.3	98.0	89.2	26.9	75.0	83.1
本文	85.5	99.5	97.9	63.2	83.2	88.6

网络	飞机	操场	立交桥	油罐
Faster R-CNN	78.7	92.2	81.2	87.6
GRoIE	79.1	90.3	81.0	87.5
HRNet	78.5	89.8	81.4	86.9
ResNet-SB	60.3	93.1	81.4	74.8
RetinaNet	59.0	59.3	30.1	57.1
NAS-FCOS	66.3	86.8	72.6	83.5
本文	79.9	90.0	85.2	87.9