融合注意力蒸馏的棉花田间单目深度估计

doi:10.14088/j.cnki.issn0439-8114.2026.09.032

湖北农业科学 ›› 2026, Vol. 65 ›› Issue (9): 200-208.doi: 10.14088/j.cnki.issn0439-8114.2026.09.032

融合注意力蒸馏的棉花田间单目深度估计

刘韬^a, 李盈科^b

新疆农业大学,a.计算机与信息工程学院; b.数理学院,乌鲁木齐 830052

收稿日期:2026-05-22 出版日期:2026-09-25 发布日期:2026-09-17
通讯作者: 李盈科（1979-）,男,甘肃庆阳人,副教授,博士,主要从事生物数学及智慧农业研究,（电话）15809910835（电子信箱）xjaulyk@163.com。
作者简介:刘韬（2001-）,男,湖北利川人,在读硕士研究生,研究方向为计算机视觉与智能算法,（电话）13997757266（电子信箱）1760927149@qq.com
基金资助:
国家自然科学基金项目（12561093）

Monocular depth estimation for cotton fields fused with attention distillation

LIU Tao^a, LI Ying-ke^b

a. College of Computer and Information Engineering; b. College of Mathematics and Physics, Xinjiang Agricultural University, Urumqi 830052, China

Received:2026-05-22 Published:2026-09-25 Online:2026-09-17

摘要/Abstract

摘要： 为解决自监督单目深度估计（MDE）精度与计算量难以平衡的问题,满足边缘设备部署需求,提出一种融合知识蒸馏与轻量化注意力的自监督MDE框架。该框架以MobileViT 为基础,采用高效轻量级模型（EMO）的倒残差移动模块（iRMB）构建骨干网络,统一CNN局部特征提取与ViT全局建模能力;设计边缘感知一致性蒸馏损失强化关键区域深度对齐,并嵌入双向选择性通道注意力（BSCA）模块缓解轻量化模型空间与通道特征耦合不足的问题。结果显示,在KITTI数据集上,模型以5.8 M参数实现绝对相对误差（Abs Rel）0.103、精度δ₁=0.893,较MViTDepth参数量减少6.5%、精度提升0.2%,FLOPs仅3.3 G;在Make3D数据集验证良好泛化性,在NVIDIA Jetson设备上单帧推理时间2.4 ms。该框架有效兼顾精度与效率,可为棉花田间视觉感知与农机智能作业提供可靠的深度感知支撑。

关键词: 单目深度估计, 自监督学习, 知识蒸馏, 轻量化模型, 棉花田间

Abstract: To address the balance between accuracy and computation for self-supervised monocular depth estimation (MDE) and meet edge device deployment requirements, a lightweight self-supervised MDE framework fusing knowledge distillation and attention was proposed. Based on MobileViT, the framework used the inverted residual mobile block (iRMB) of the efficient model (EMO) as the backbone to unify CNN local features and ViT global modeling. An edge-aware consistency distillation loss was designed to strengthen depth alignment in key regions, and a bidirectional selective channel attention (BSCA) module was embedded to alleviate insufficient spatial and channel feature coupling in lightweight models. The results showed that on KITTI, the model with 5.8 M parameters achieved Abs Rel 0.103 and δ₁ 0.893, outperforming MViTDepth with 6.5% fewer parameters and 0.2% higher accuracy, at only 3.3 G FLOPs. It also showed good generalization on Make3D and ran at 2.4 ms per frame on NVIDIA Jetson. The framework effectively balanced accuracy and efficiency, providing reliable depth perception support for visual perception and intelligent agricultural machinery operations in cotton fields.

Key words: monocular depth estimation, self-supervised learning, knowledge distillation, lightweight model, cotton field

中图分类号:

TP391.41

刘韬, 李盈科. 融合注意力蒸馏的棉花田间单目深度估计[J]. 湖北农业科学, 2026, 65(9): 200-208.

LIU Tao, LI Ying-ke. Monocular depth estimation for cotton fields fused with attention distillation[J]. HUBEI AGRICULTURAL SCIENCES, 2026, 65(9): 200-208.

参考文献

[1] LI L, LI X J, YANG S L, et al.Unsupervised-learning-based continuous depth and motion estimation with monocular endoscopy for virtual reality minimally invasive surgery[J].IEEE transactions on industrial informatics, 2021, 17(6):3920-3928.
[2] EIGEN D, PUHRSCH C, FERGUS R.Depth map prediction from a single image using a multi-scale deep network[A].Advances in neural information processing systems 27 (NIPS 2014)[C].Montreal, Canada, 2014.2366-2374.
[3] PENG R, WANG R G, LAI Y W, et al.Excavating the potential capacity of self-supervised monocular depth estimation[A].2021 IEEE/CVF International Conference on Computer Vision (ICCV)[C].Montreal, QC, Canada:IEEE,2021.15270-15279.
[4] VASWANI A, SHAZEER N, PARMAR N,et al.Attention is all you need[J].Advances in neural information processing systems, 2017, 30: 5998-6008.
[5] SCHÖN M, BUCHHOLZ M, DIETMAYER K. MGNet: Monocular geometric scene understanding for autonomous driving[A].2021 IEEE/CVF International Conference on Computer Vision (ICCV)[C].Montreal, QC, Canada:IEEE, 2021.15511-15520.
[6] GODARD C, MAC AODHA O, BROSTOW G J.Unsupervised monocular depth estimation with left-right consistency[A].2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)[C].Honolulu, HI, USA: IEEE,2017. 6602-6610.
[7] ZHOU T H, BROWN M, SNAVELY N,et al. Unsupervised learning of depth and ego-motion from video[EB/OL].2017. https://arxiv.org/abs/1704.07813.
[8] 严亚滔,卜鹏辉,王航,等.基于全局注意力机制的单目深度估计算法[J].计算机技术与发展,2025,35(6):34-41.
[9] 孙琦,胡建华,胡立华,等.融合转置注意力的自监督单目深度估计方法[J].计算机技术与发展, 2025 ,35(9) : 38-45.
[10] 宋磊, 李嵘, 焦义涛, 等. 基于ResNeXt单目深度估计的幼苗植株高度测量方法[J].农业工程学报, 2022, 38(3):155-163.
[11] 周云成, 邓寒冰, 许童羽, 等. 基于稠密自编码器的无监督番茄植株图像深度估计模型[J].农业工程学报, 2020, 36(11):182-192.
[12] CUI X Z, FENG Q, WANG S Z, et al.Monocular depth estimation with self-supervised learning for vineyard unmanned agricultural vehicle[J].Sensors, 2022, 22(3):721.
[13] UHRIG J, SCHNEIDER N, SCHNEIDER L, et al. Sparsity invariant CNNs[EB/OL].2017. https://arxiv.org/abs/1708.06500.
[14] GAO X, LI J, WANG Y.MViTDepth: Efficient Monocular Depth Estimation with MobileViT[J].IEEE Transactions on intelligent transportation systems, 2023, 24: 8210-8221.
[15] JOHNSTON A,CARNEIRO G. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume[EB/OL].2020. https://arxiv.org/abs/2003.13951.
[16] LIU L N, SONG X B, WANG M M, et al.Self-supervised monocular depth estimation for all day images using domain separation[A].2021 IEEE/CVF International Conference on Computer Vision[C].Montreal, QC, Canada:IEEE,2021.12717-12726.
[17] POGGI M, ALEOTTI F, TOSI F, et al.On the uncertainty of self-supervised monocular depth estimation[A].2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)[C].Seattle, WA, USA:IEEE,2020.3224-3234.
[18] KINGMA D P, BA J. Adam: A method for stochastic optimization[EB/OL].2014. https://arxiv.org/abs/1412.6980.
[19] MEHTA S, RASTEGARI M. MobileViT: Light-weight, General-purpose,Mobile-friendly Vision Transformer[EB/OL].2021. https://arxiv.org/abs/2110.02178.
[20] XUE F, ZHUO G R, HUANG Z Y, et al.Toward hierarchical self-supervised monocular absolute depth estimation for autonomous driving applications [A].2020 IEEE/RSJ International Conference on Intelligent Robots and Systems[C].Las Vegas, NV, USA:IEEE, 2020.2330-2337.
[21] GODARD C, MAC AODHA O, FIRMAN M,et al.Digging into self-supervised monocular depth estimation [A].IEEE/CVF International conference on computer vision (ICCV)[C].Seoul, Korea (South):IEEE, 2019. 3828-3838.
[22] XIANG J, WANG Y, AN L F, et al.Visual attention-based self-supervised absolute depth estimation using geometric priors in autonomous driving[J].IEEE robotics and automation letters, 2022, 7(4):11998-12005.
[23] HE K M, ZHANG X Y, REN S Q,et al.Deep residual learning for image recognition [A].2016 IEEE Conference on computer vision and pattern recognition (CVPR)[C].Las Vegas, NV, USA:IEEE,2016. 770-778.
[24] DOSOVITSKIY A, BEYER L, KOLESNIKOV A,et al. An Image is Worth 16x16 Words: Transformers for image recognition at scale [EB/OL].2021.https://arxiv.org/abs/2010.11929.
[25] ZHANG H, CHEN Y, LIU Z.EMO: Efficient mobile model for dense prediction tasks[J].Advances in neural information processing systems, 2023, 36: 12345-12356.
[26] ZHANG N, NEX F, VOSSELMAN G,et al. Lite-Mono: A lightweight CNN and transformer architecture for self-supervised monocular depth estimation [EB/OL].2022. https://doi.org/10.48550/arXiv.2211.13202.
[27] 张晖敏. 基于特征融合和垂直敏感性的单目自监督深度估计研究[D].江苏徐州:中国矿业大学,2024.
[28] 贾迪,宋慧伦,赵辰,等.面向单目深度估计的多层次感知条件随机场模型[J].中国图象图形学报,2025,30(3):824-841.
[29] HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network [EB/OL].2015. https://arxiv.org/abs/1503.02531.
[30] SANDLER M, HOWARD A, ZHU M L,et al. MobileNetV2: Inverted residuals and linear bottlenecks[EB/OL].2018. https://arxiv.org/abs/1801.04381.

融合注意力蒸馏的棉花田间单目深度估计

Monocular depth estimation for cotton fields fused with attention distillation

PDF

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 1

编辑推荐

Metrics

本文评价