联合正交特征融合与大核可分离注意力的道路分割算法

doi:10.3785/j.issn.1008-973X.2026.09.011

浙江大学学报(工学版)

2026, Vol. 60

Issue (9): 1942-1952 DOI: 10.3785/j.issn.1008-973X.2026.09.011

计算机技术、自动控制技术

联合正交特征融合与大核可分离注意力的道路分割算法

汤毅杰1(

),钟铭恩1,*(

),袁彬淦2,范康3,谭佳威3,林志强1

1. 厦门理工学院机械与汽车工程学院，福建省客车先进设计与制造重点实验室，福建厦门 361024
2. 清华大学工程物理系，北京 100084
3. 厦门大学航空航天学院，福建厦门 361005

Road segmentation algorithm based on joint orthogonal feature fusion and large kernel separable attention

Yijie TANG1(

),Mingen ZHONG1,*(

),Bingan YUAN2,Kang FAN3,Jiawei TAN3,Zhiqiang LIN1

1. Key Laboratory of Advanced Design and Manufacturing for Buses in Fujian Province, School of Mechanical and Automotive Engineering, Xiamen University of Technology, Xiamen 361024, China
2. Department of Engineering Physics, Tsinghua University, Be?ing 100084, China
3. School of Aerospace Engineering, Xiamen University, Xiamen 361005, China

全文: PDF(1960 KB) HTML

摘要：

针对航拍图像中的道路结构模糊、边缘不连贯等特点，提出轻量级道路分割算法ARSNet. 设计正交特征融合模块（OFFM），实现不同方向图像特征间的交互与增强. 构建上下文特征引导聚合模块，提升算法在道路边缘拓扑结构提取时的连贯性. 在特征空间池化操作中引入大核可分离注意力，加强对各种尺度和方向上的道路特征的协同感知能力. 基于神经网络权重幅值进行层自适应稀疏剪枝，将算法参数量降低到22.5×10⁶，FLOPs控制在14.8×10⁹. 在公开数据集Massachusetts与DeepGlobe上的实验结果表明，所提算法的道路分割精确度分别为78.12%和79.78%，召回率为79.63%和81.63%，像素交并比为64.85%和67.88%，平均路径长度相似度达到74.95%和75.53%. ARSNet在保持较高精度的前提下，推理速度达到178帧/s，整体性能优于所对比的主流算法.

关键词： 航拍图像; 道路分割; 正交特征融合; 上下文特征引导聚合; 大核可分离注意力; 模型剪枝

Abstract:

A lightweight road segmentation algorithm for aerial imagery, termed ARSNet, was proposed to address the issue that roads in aerial images often suffer from structural ambiguity and discontinuous edges. An orthogonal feature fusion module (OFFM) was designed to achieve interaction and enhancement among features in different orientations. A context-guided feature aggregation module (C2f-CGA) was introduced to improve the continuity of road boundary and topological structure extraction. Large-kernel separable attention was introduced into the feature spatial pooling module in order to strengthen the collaborative perception of road features across various scales and directions. A layer-adaptive sparse pruning strategy based on weight magnitude was employed to reduce the number of parameters to 22.5×10⁶ with FLOPs controlled at 14.8×10⁹. Experiments conducted on the public Massachusetts and DeepGlobe datasets showed that ARSNet achieved precision of 78.12% and 79.78%, recall of 79.63% and 81.63%, intersection-over-union of 64.85% and 67.88%, and average path length similarity of 74.95% and 75.53%, respectively. ARSNet achieved an inference speed of 178 frames per second while maintaining high accuracy, consistently outperforming a series of state-of-the-art algorithms.

Key words: aerial imagery road segmentation orthogonal feature fusion context feature-guided aggregation large kernel separable attention model pruning

收稿日期: 2025-08-06 出版日期: 2026-07-20

CLC:

TP 391

基金资助: 福建省自然科学基金资助项目（2023J011439）.

通讯作者: 钟铭恩 E-mail: 2421011030@stu.xmut.edu.cn;zhongmingen@xmut.edu.cn

作者简介: 汤毅杰（2000—），男，硕士生，从事机器视觉和智慧交通研究. orcid.org/0009-0008-6120-4484. E-mail：2421011030@stu.xmut.edu.cn

	服务
	把本文推荐给朋友
	加入引用管理器
	E-mail Alert
	作者相关文章
	汤毅杰
	钟铭恩
	袁彬淦
	范康
	谭佳威
	林志强

引用本文:

汤毅杰,钟铭恩,袁彬淦,范康,谭佳威,林志强. 联合正交特征融合与大核可分离注意力的道路分割算法[J]. 浙江大学学报(工学版), 2026, 60(9): 1942-1952.

Yijie TANG,Mingen ZHONG,Bingan YUAN,Kang FAN,Jiawei TAN,Zhiqiang LIN. Road segmentation algorithm based on joint orthogonal feature fusion and large kernel separable attention. Journal of ZheJiang University (Engineering Science), 2026, 60(9): 1942-1952.

链接本文:

https://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2026.09.011 或 https://www.zjujournals.com/eng/CN/Y2026/V60/I9/1942

图 1 ARSNet整体结构

图 2 正交特征融合模块结构图

图 3 上下文特征引导聚合模块结构图

图 4 大核可分离注意力空间池化模块结构图

表 1 不同算法的性能对比结果

图 5 不同算法在Massachusetts数据集上的分割结果对比

图 6 不同算法在DeepGlobe数据集上的分割结果对比

表 2 消融实验结果

图 7 在Massachusetts数据集上的模块消融实验可视化对比

图 8 在DeepGlobe数据集上的模块消融实验可视化对比

表 3 不同剪枝方法性能对比结果

表 4 LAMP剪枝方法下剪枝率与IoU、APLS关系对比

1	QI Y, HE Y, QI X, et al. Dynamic snake convolution based on topological geometric constraints for tubu-lar structure segmentation [C]// IEEE/CVF International Conference on Computer Vision. Paris: IEEE, 2023: 6070-6079.
2	王艳军, 唐徐超, 王成, 等基于道路拓扑关联特征的城乡道路面精细提取网络[J]. 测绘学报, 2025, 54 (1): 75- 89 WANG Yanjun, TANG Xuchao, WANG Cheng, et al Urban and rural road surface extraction network based on topological correlation features[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54 (1): 75- 89
3	DEY M, P S P, AITHAL B H. UnetEdge: a transfer learning-based framework for road feature segmentation from high-resolution remote sensing images[J]. Remote Sensing Applications: Society and Environment, 2024, 34: 101160 doi: 10.1016/j.rsase.2024.101160
4	LI X, YANG S, MENG F, et al LCMorph: exploiting frequency cues and morphological perception for low-contrast road extraction in remote sensing images[J]. Remote Sensing, 2025, 17 (2): 257 doi: 10.3390/rs17020257
5	WANG Y, SEO J, JEON T NL-LinkNet: toward lighter but more accurate road extraction with nonlocal operations[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19: 1- 5
6	YANG M, YUAN Y, LIU G SDUNet: road extraction via spatial enhanced and densely connected UNet[J]. Pattern Recognition, 2022, 126: 108549 doi: 10.1016/j.patcog.2022.108549
7	LEE J, PARK S, MO S, et al. Layer-adaptive sparsity for the magnitude-based pruning [EB/OL]. (2020-11-15)[2025-06-05]. https://arxiv.org/pdf/2010.07611.
8	YUAN T, HU B REU-Net: a remote sensing image building segmentation network based on residual structure and the edge enhancement attention module[J]. Applied Sciences, 2025, 15 (6): 3026 doi: 10.3390/app15063026
9	SUREKHA J N, NEHA G SAR U-Net: spatial attention residual U-Net structure for water body segmentation from remote sensing satellite images[J]. Multimedia Tools and Applications, 2023, 83 (15): 44425- 44454 doi: 10.1007/s11042-023-16965-8
10	TERVEN J, CORDOVA-ESPARZA D M, ROMERO-GONZALEZ J A A comprehensive review of Yolo architectures in computer vision: from Yolov1 to Yolov8 and Yolo-NAS[J]. Machine Learning and Knowledge Extraction, 2023, 5 (4): 1680- 1716 doi: 10.3390/make5040083
11	范康, 钟铭恩, 谭佳威, 等联合语义分割和深度估计的交通场景感知算法[J]. 浙江大学学报: 工学版, 2024, 58 (4): 684- 695 FAN Kang, ZHONG Mingen, TAN Jiawei, et al Traffic scene perception algorithm based on joint semantic segmentation and depth estimation[J]. Journal of Zhejiang University: Engineering Science, 2024, 58 (4): 684- 695
12	JIN Y H, GAN H Y, LI J, et al ResEff-YOLO: accuracy enhancement of YOLOv8 through integration of ResNet, SPPF, and EfficientHead modules[J]. International Core Journal of Engineering, 2025, 11 (2): 23- 34
13	LAU K W, PO L M, REHMAN Y A U Large separable kernel attention: rethinking the large kernel attention design in CNN[J]. Expert Systems with Applications, 2024, 236: 121352 doi: 10.1016/j.eswa.2023.121352
14	ZHANG X L, LU L, LUO H Y, et al Improvement in pavement defect scenarios using an improved YOLOv10 with ECA attention, RefConv and WIoU[J]. World Electric Vehicle Journal, 2025, 16 (6): 328 doi: 10.3390/wevj16060328
15	MNIH V. Machine Learning for aerial image labeling [D]. Toronto Q12: University of Toronto, 2013.
16	DEMIR I, KOPERSKI K, LINDENBAUM D, et al. Deepglobe 2018: A challenge to parse the earth through satellite images [C]// IEEE Conference on Computer Vision and Pattern Recognition Workshops. Salt Lake City: IEEE, 2018: 172–181.
17	VAN ETTEN A, LINDENBAUM D, BACASTOW T M. SpaceNet: a remote sensing dataset and challenge series [EB/OL]. (2019-07-15)[2025-06-05]. https://arxiv.org/abs/1807.01232.
18	RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation [C]// Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2015: 234-241.
19	BADRINARAYANAN V, KENDALL A, CIPOLLA R, et al SegNet: a deep convolutional encoder-decoder architecture for image segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39 (12): 2481- 2495
20	CHEN L, ZHU Y, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation [C]// European Conference on Computer Vision. Munich: IEEE, 2018: 801–818.
21	ZHOU L, ZHANG C, WU M, et al. D-linknet: Linknet with pretrained encoder and dilated convolution for high resolution satellite imagery road extraction [C]// IEEE Conference on Computer Vision and Pattern Recognition Workshop. Salt Lake City: IEEE, 2018: 182–186.
22	CHEN H, LI Z, WU J, et al SemiRoadExNet: a semi-supervised network for road extraction from remote sensing imagery via adversarial learning[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2023, 198: 169- 183 doi: 10.1016/j.isprsjprs.2023.03.012
23	YANG R, ZHONG Y, LIU Y, et al Occlusion-aware road extraction network for high-resolution remote sensing imagery[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5619316
24	WANG H, BAI L, XUE D, et al FRCFNet: feature reassembly and context information fusion network for road extraction[J]. IEEE Geoscience and Remote Sensing Letters, 2024, 21: 2502805
25	XIONG Y, LI L, YUAN D, et al CFRNet: road extraction in remote sensing images based on cascade fusion network[J]. IEEE Geoscience and Remote Sensing Letters, 2024, 21: 6011705
26	ZHU E Z, CHEN Z, WANG D K, et al UNetMamba: an efficient UNet-Like Mamba for semantic segmentation of high-resolution remote sensing images[J]. IEEE Geoscience and Remote Sensing Letters, 2025, 22: 6001205
27	CHEN S H, YU L J, LIU Z, et al. LightFormer: a lightweight and efficient decoder for remote sensing image segmentation [EB/OL]. (2025-04-15)[2025-06-05]. https://arxiv.org/pdf/2504.10834.
28	CARRERAS J Celiac disease Deep Learning image classification using convolutional neural networks[J]. Journal of Imaging, 2024, 10 (8): 200 doi: 10.3390/jimaging10080200
29	刘洲峰, 吴文涛, 李环宇, 等聚类和群智能优化算法的自动剪枝方法[J]. 计算机工程与应用, 2025, 61 (11): 204- 215 LIU Zhoufeng, WU Wentao, LI Huanyu, et al Automatic channel pruning method based on clustering and swarm intelligence optimization algorithm[J]. Computer Engineering and Applications, 2025, 61 (11): 204- 215
30	FANG G, MA X, SONG M, et al. Depgraph: towards any structural pruning [C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 16091-16101.
31	YU S X, YAO Z W, GHOLAMI A, et al. Hessian-aware pruning and optimal neural implant [C]// IEEE/CVF Winter Conference on Applications of Computer Vision. Waikoloa, HI: IEEE, 2022: 3665-3676.
32	MOLCHANOV P, MALLYA A, TYREE S, et al. Importance estimation for neural network pruning [C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2019: 11264-11272.
33	SONG H X, XIE J P, DUAN Y Y, et al CMKD-Net: a cross-modal knowledge distillation method for remote sensing image classification[J]. Advances in Space Research, 2025, 75 (12): 8515- 8534 doi: 10.1016/j.asr.2025.04.009

[1]	李珺,丁彬彬,史维娟,杨琳. 基于多尺度特征聚合的航拍图像检测算法[J]. 浙江大学学报(工学版), 2026, 60(9): 1890-1900.
[2]	邬开俊,郑云琦,魏鼎,袁海翔. 基于YOLOv8s的轻量化航拍图像小目标检测算法[J]. 浙江大学学报(工学版), 2026, 60(9): 1912-1923.
[3]	夏杰锋,唐武勤,杨强. 光伏航拍红外图像的热斑自动检测方法[J]. 浙江大学学报(工学版), 2022, 56(8): 1640-1647.
[4]	陈彤,郭剑锋,韩心中,谢学立,席建祥. 基于生成对抗模型的可见光-红外图像匹配方法[J]. 浙江大学学报(工学版), 2022, 56(1): 63-74.
[5]	刘加海白洪欢黄微凹. 基于彩色和边缘信息融合的道路分割算法[J]. J4, 2006, 40(1): 29-32.

Viewed

Full text

Abstract

Cited

Shared

Discussed