KD-NeRF：知识蒸馏增强的果树神经辐射场重建方法

doi:10.3969/j.issn.1006-2475.2025.12.010

摘要/Abstract

摘要： 摘要：精准农业对果树三维表型精确重建需求迫切。传统方法存在建模效率低、计算成本高的问题，同时二维分析受限于遮挡和深度缺失。神经辐射场（NeRF）虽表现优异，但训练耗时且资源密集，学习果树复杂结构面临挑战。本文提出一种融合知识蒸馏增强的NeRF果树三维重建方法(KD-NeRF)。该方法构建基于二维多视角图像的深度学习框架；采用SIREN周期性激活函数，通过隐式正则化增强高频细节捕获和空间连续性；设计教师-学生知识蒸馏机制，将深层表征传递至轻量级网络，提升复杂结构学习效能。在不同果树数据集上进行实验的结果表明，相较NeRF、KD-NeRF的PSNR提升8%，训练速度提升14倍~16倍。消融实验结果表明：教师-学生架构的PSNR、SSIM等指标有明显提升；SIREN增强了高频细节表征；两者协同产生了超越单一累加的性能增益。KD-NeRF解决了NeRF在果树重建中的效率瓶颈，可为果树表型分析、生长监测和智能育种提供良好技术支撑，推动了精准农业和智慧果园建设。

关键词: 关键词：深度学习, 农业技术, 三维重建, 神经辐射场, 知识蒸馏

Abstract: Abstract: Precision agriculture urgently demands accurate three-dimensional phenotypic reconstruction of fruit trees. Traditional methods suffer from low modeling efficiency and high computational costs, while two-dimensional analysis is constrained by occlusion and a lack of depth information. Although Neural Radiance Fields (NeRF) demonstrate excellent performance, they are time-consuming and resource-intensive to train, facing significant challenges in learning complex fruit tree structures. This study proposes a knowledge distillation-enhanced NeRF method for three-dimensional fruit tree reconstruction (KD-NeRF). We construct a deep learning framework based on two-dimensional multi-view images and employ SIREN periodic activation functions to enhance high-frequency detail capture and spatial continuity through implicit regularization. A teacher-student knowledge distillation mechanism is designed to transfer deep representations to lightweight networks, improving the learning efficiency for complex structures. Experiments on diverse fruit tree datasets demonstrate that compared to NeRF, KD-NeRF achieves an 8% improvement in PSNR with 14 times ~16 times acceleration in training speed. The results of the ablation experiment confirm that the teacher-student architecture significantly improves PSNR, SSIM, and other metrics; SIREN enhances high-frequency detail representation; and their synergy produces performance gains exceeding simple additive effects. KD-NeRF addresses the efficiency bottleneck of NeRF in fruit tree reconstruction, providing efficient technical support for fruit tree phenotypic analysis, growth monitoring, and intelligent breeding, thereby advancing precision agriculture and smart orchard construction.

Key words: Key words: deep learning, agricultural technology, 3D reconstruction, neural radiance fields, knowledge distillation

中图分类号:

中图分类号：TP391

汤辉1, 刘鑫2, 胡必伟1, 刘帆1. KD-NeRF：知识蒸馏增强的果树神经辐射场重建方法[J]. 计算机与现代化, 2025, 0(12): 66-73.

TANG Hui1, LIU Xin2, HU Biwei1, LIU Fan1. KD-NeRF: Knowledge Distillation-enhanced Neural Radiance Field Reconstruction Method for Fruit Trees[J]. Computer and Modernization, 2025, 0(12): 66-73.

参考文献

[1] YI W L, XIA S K, KUZMIN S, et al. RTFVE-YOLOv9: Real-time fruit volume estimation model integrating YOLOv9 and binocular stereo vision[J]. Computers and Electronics in Agriculture, 2025,236:110401.
[2] YI W L, ZHANG X S, DAI S M, et al. MV-SSRP: Machine vision approach for stress-strain measurement in rice plants[J]. Agronomy, 2024,14(7):1443.
[3] YI W L, XIA S K, KUZMIN S, et al. YOLOv7-KDT: An ensemble model for pomelo counting in complex environment[J]. Computers and Electronics in Agriculture, 2024,227:109469.
[4] LIN G Y, YANG L, ZHANG C Y, et al. Patch-Grid: An efficient and feature-preserving neural implicit surface representation[J]. ACM Transactions on Graphics, 2025,44(2):1-21.
[5] RELLA M E, CHHATKULI A, KONUKOGLU E, et al. Neural vector fields for implicit surface representation and inference[J]. International Journal of Computer Vision, 2025,133(4):1855-1878.
[6] MILDENHALL B, SRINIVASAN P P, TANCIK M, et al. NeRF: Representing scenes as neural radiance fields for view synthesis[J]. Communications of the ACM, 2021,65(1):99-106.
[7] LIAO Y, DI Y D, ZHOU H, et al. A survey on neural radiance fields[J]. ACM Computing Surveys, 2025,58(2):1-33.
[8] HU K W, YING W, PAN Y Q, et al. High-fidelity 3D reconstruction of plants using neural radiance fields[J]. Computers and Electronics in Agriculture, 2024,220:108848.
[9] ZHANG Y, WU Y Q, TONG K, et al. Review of visual simultaneous localization and mapping based on deep learning[J]. Remote Sensing, 2023,15(11):2740.
[10] WANG J, BAO W D, SUN L C, et al. Private model compression via knowledge distillation[C]// Proceedings of the AAAI Conference on Artificial Intelligence. AAAI， 2019:1190-1197.
[11] PARK Y, HWANG J, JEONG S M, et al. MSPKD: Multi spatial projectors for knowledge distillation in semantic segmentation[J]. Machine Vision and Applications, 2025,36(4). DOI：10.1007/s00138-025-01721-9.
[12] MALIK A, MIRDEHGHAN P, NOUSIAS S, et al. Transient neural radiance fields for lidar view synthesis and 3D reconstruction[C]// Proceedings of the 37th International Conference on Neural Information Processing Systems. ACM, 2023:71569-71581.
[13] KAJIYA J T, VON HERZEN B P. Ray tracing volume densities[J]. ACM SIGGRAPH Computer Graphics, 1984,18(3):165-174.
[14] GUO Z Y, ZHOU W G, WANG M, et al. HandNeRF++: Modeling animatable interacting hands with neural radiance fields[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025,47(9):8102-8116.
[15] CHRISTODOULIDES A, TAM G K L, CLARKE J, et al. Survey on 3D reconstruction techniques: Large-scale urban city reconstruction and requirements[J]. IEEE Transactions on Visualization and Computer Graphics, 2025,31(10):9343-9367.
[16] ZHANG K, RIEGLER G, SNAVELY N, et al. Nerf++: Analyzing and improving neural radiance fields[J]. arXiv preprint arXiv:2010.07492, 2020.
[17] SITZMANN V, THIES J, HEIDE F, et al. DeepVoxels: Learning persistent 3D feature embeddings[C]// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2019:2437-2446.
[18] NEFF T, STADLBAUER P, PARGER M, et al. DONeRF: Towards real-time rendering of compact neural radiance fields using depth oracle networks[J]. Computer Graphics Forum, 2021,40(4):45-59.
[19] SRINIVASAN P P, DENG B, ZHANG X, et al. NeRV: Neural reflectance and visibility fields for relighting and view synthesis[C]// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2021:7495-7504.
[20] LIU Z, ZHU H, ZHANG Q, et al. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions[C]// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024:2713-2722.
[21] HUANG Z C, BAI S J, KOLTER J Z. (Implicit)2: Implicit layers for implicit representations[C]// Proceedings of 35th Annual Conference on Neural Information Processing Systems. ACM, 2021:9639-9650.
[22] CHAN E R, MONTEIRO M, KELLNHOFER P, et al. pi-GAN: Periodic implicit generative adversarial networks for 3D-aware image synthesis[C]// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2021:5799-5809.
[23] XIA Y T, TANG H, TIMOFTE R, et al. SiNeRF: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction[J]. arXiv preprint arXiv:2210.04553,
2022.
[24] CHEN Z Q, FUNKHOUSER T, HEDMAN P, et al. MobileNeRF: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures[C]// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2023:16569-16578.
[25] GARBIN S J, KOWALSKI M, JOHNSON M, et al. FastNeRF: High-fidelity neural rendering at 200fps[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. IEEE, 2021:14346-14355.
[26] HEDMAN P, SRINIVASAN P P, MILDENHALL B, et al. Baking neural radiance fields for real-time view synthesis[C]// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. IEEE, 2021:5875-5884.
[27] KURZ A, NEFF T, LV Z Y, et al. AdaNeRF: Adaptive sampling for real-time rendering of neural radiance fields[C]// Proceedings of the 17th European Conference on Computer Vision. Springer, 2022:254-270.
[28] PIALA M, CLARK R. TermiNeRF: Ray termination prediction for efficient neural rendering[C]// Proceedings of the 2021 International Conference on 3D Vision (3DV). IEEE, 2021:1106-1114.
[29] ATTAL B, HUANG J B, ZOLLHÖFER M, et al. Learning neural light fields with ray-space embedding[C]// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2022:19819-19829.
[30] LIU L J, GU J T, LIN Z K, et al. Neural sparse voxel fields[C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. ACM, 2020:15651
-15663.
[31] YU A, LI R, TANCIK M, et al. PlenOctrees for real-time rendering of neural radiance fields[C]// Proceedings of the 2021 IEEE/CVF Computer Vision and Pattern Recognition. IEEE, 2021:5752-5761.
[32] MÜLLER T, EVANS A, SCHIED C, et al. Instant neural graphics primitives with a multiresolution hash encoding[J]. ACM Transactions on Graphics (TOG), 2022,41(4). DOI: 10.1145/3528223.3530127.
[33] REBAIN D, JIANG W, YAZDANI S, et al. DeRF: Decomposed radiance fields[C]// Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2021:14153-14161.
[34] WANG Z R, WU S Z, XIE W D, et al. NeRF--: Neural radiance fields without known camera parameters[J]. arXiv preprint arXiv:2102.07064, 2021.
[35] GOU J P, YU B S, MAYBANK S J, et al. Knowledge distillation: A survey[J]. International Journal of Computer Vision, 2021,129(6):1789-1819.
[36] MILDENHALL B, SRINIVASAN P P, ORTIZ-CAYON R, et al. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines[J]. ACM Transactions on Graphics (ToG), 2019,38(4). DOI: 10.1145/3306346.3322980.
[37] SCHONBERGER J L, FRAHM J M. Structure-from-motion revisited[C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2016:4104-4113.
[38] SARA U, AKTER M, UDDIN M S. Image quality assessment through FSIM, SSIM, MSE and PSNR—A comparative study[J]. Journal of Computer and Communications, 2019,7(3):8-18.
[39] WANG Z, BOVIK A C, SHEIKH H R, et al. Image quality assessment: From error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004,13(4):600-612.
[40] ZHANG R, ISOLA P, EFROS A A, et al. The unreasonable effectiveness of deep features as a perceptual metric[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2018:586-595.
[41] ROSS A, WILLSON V L. Paired samples T-test[M]// Basic and advanced statistical tests: Writing results sections and creating tables and figures. Rotterdam: SensePublishers, 2017:17-19.
[42] CHEN A, XU Z X, GEIGER A, et al. TensoRF: Tensorial radiance fields[C]// Proceedings of the 2022 European Conference on Computer Vision. Springer, 2022:333-350.
[43] HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[J]. arXiv preprint arXiv:1503.02531, 2015.
[44] AHN S X, HU S, DAMIANOU A, et al. Variational information distillation for knowledge transfer[C]// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2019:9163-9171.
[45] CHEN G B, CHOI W, YU X, et al. Learning efficient object detection models with knowledge distillation[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. ACM, 2017:742-751.
[46] ZHOU B L, ZHAO H, PUIG X, et al. Semantic understanding of scenes through the ADE20K dataset[J]. International Journal of Computer Vision, 2019,127(3):302-321.