融合扩散模型和流形梯度约束的无人机图像拼接方法

doi:10.7523/j.ucas.2024.061

摘要/Abstract

摘要：

无人机图像拼接是无人机遥感应用的必要前置步骤，但拼接结果中通常会存在大块不规则边界和多个拼接缝，影响后续分析和应用。现有方法通常不能同时解决这2个问题，需要对图像拼接结果进行多次处理，造成多次误差传递。针对这一问题，提出一种基于去噪离散扩散模型（DDPM）的无人机图像拼接补全方法，该方法对不规则边界和拼接缝统一设计掩码，利用带有流形梯度先验约束的扩散模型补全掩码区域的图像，可同时消除不规则边界与拼接缝，改进拼接结果的质量。利用不同场景数据开展对比实验，结果表明该方法有效消除了拼接结果中的不规则边界与拼接缝，修复后从局部到全局图像质量评价（PaQ-2-PiQ）与多尺度图像质量转换评价模型（MUSIQ）得分分别提升4.36%和15.37%，不规则边界处的结构相似性（SSIM）和峰值信噪比（PSNR）分别提升20.22%与33.69%。与SOTA方法和其他经典的图像拼接算法的对比实验结果表明，该方法在主观和客观2种质量评价方式下都优于其他方法，具有良好的鲁棒性和泛化性，可广泛应用于无人机图像拼接场景。

关键词: 无人机图像, 扩散模型, 图像拼接, 不规则边界, 拼接缝, 流形梯度约束

Abstract:

Image stitching is a crucial prerequisite step for unmanned aerial vehicle （UAV） remote sensing applications， while the stitched images using most of the current image stitching methods often suffer from large irregular boundaries and multiple stitching seams， which can seriously affect subsequent analysis and applications. Existing improved methods typically cannot simultaneously address these two issues， and integrating the two types of methods in sequence is a straightforward way to solve the two problems， while this often can not obtain satisfactory performance because of the inevitable error propagation problem. This paper proposes an inpainting method for the UAV image stitching task based on the denoising diffusion probability model （DDPM）. The method uniformly designs masks for irregular boundaries and stitching seams， and a diffusion model is then utilized with manifold gradient prior constraints to complete the masked regions. By doing so， both irregular boundaries and stitching seams are simultaneously eliminated， thereby improving the quality of the stitching results. Comparative experiments are conducted using four datasets established for different scenarios. The experimental results demonstrated the efficacy of the proposed method in effectively eliminating irregular boundaries and seams in the image stitching. Moreover， from patches to pictures quality （PaQ-2-PiQ） and multi-scale image quality （MUSIQ） scores increased by 4.36% and 15.37%， respectively. Furthermore， at the locations of irregular boundaries， the structural similarity （SSIM） and peak signal-to-noise ratio （PSNR） values improved by 20.22% and 33.69%， respectively. Compared with state-of-the-art methods and other conventional image stitching algorithms， the proposed method performs better in both subjective and objective quality metric scores， has good robustness and generalization， and can be widely applied to UAV image stitching scenarios.

Key words: UAV image, diffusion model, image stitching, irregular boundaries, stitching seams, manifold gradient constraint

中图分类号:

TP751.1

王杰, 罗永曦, 陈俊, 吴业炜. 融合扩散模型和流形梯度约束的无人机图像拼接方法[J]. 中国科学院大学学报, 2026, 43(2): 252-264.

Jie WANG, Yongxi LUO, Jun CHEN, Yewei WU. UAV image stitching method based on diffusion model and manifold gradient constraint[J]. Journal of University of Chinese Academy of Sciences, 2026, 43(2): 252-264.

图/表 15

图1 无人机图像拼接结果中的不规则边界与拼接缝

Fig.1 Irregular boundaries and stitching seams in the stitched images of the unmanned aerial vehicle

图2 所提方法的总体流程图MGC的细节： ① 逆扩散的初始采样； ② 使用流形梯度约束执行反向迭代； ③ 取掩码的正交补； ④ 对带掩码的图像进行扩散。

Fig.2 Overall diagram of the proposed method

图3 带时间嵌入向量的U-Net

Fig.3 U-Net with time embedding vectors

图4 本文方法拼接结果修复前后对比

Fig.4 Comparison of results before and after restoration using the proposed method

图5 本文方法拼接结果修复前后细节对比

Fig.5 Comparison of details before and after restoration using the proposed method

表1 本文方法修复前后图像的客观质量比较

Table 1 Objective quality comparison of images before and after restoration using the proposed method

图像

PaQ-2-Piq

MUSIQ

修复前

修复后

0.734

0.766

0.631

0.728

图6 不规则边界的修复值与真实值的对比

Fig.6 Comparison between repaired values and ground truth for irregular boundaries

表2 本文方法修复前后与参考值相似度度量

Table 2 Measurement of similarity with the reference value between images before and after restoration using the proposed method

评价指标

SSIM

PSNR

修复前

修复后

0.361

0.434

12.724

17.011

图7 与其他矩形化方法拼接结果对比

Fig.7 Comparison of stitching results with other rectangling methods

表 3 与其他矩形化方法感知质量对比

Table 3 Comparison of perceptual quality with other rectangling methods

方法

PaQ-2-PiQ

MUSIQ

Deep Rectangling

本文方法(仅使用不规则边界掩码)

0.742

0.751

0.648

0.708

图8 与去除拼接缝方法拼接结果对比

Fig.8 Comparison of stitching results with seamless stitching methods

表 4 与去除拼接缝方法感知质量对比

Table 4 Comparison of perceptual quality with seamless stitching methods

方法

PaQ-2-PiQ

MUSIQ

UDIS++

本文方法(仅使用拼接缝掩码)

0.695

0.743

0.629

0.676

图9 两两拼接场景下的拼接结果对比

Fig.9 Comparison of stitching results in pairwise stitching scenarios

图10 两两拼接场景下拼接结果的局部细节对比

Fig.10 Local detail comparison of stitching results in pairwise stitching scenarios

表 5 两两拼接场景下的感知质量对比

Table 5 Comparison of perceptual quality in pairwise stitching scenarios

方法

PaQ-2-PiQ

MUSIQ

SPW

LPC

APAP

AANAP

本文方法

0.691

0.711

0.643

0.652

0.728

0.686

0.706

0.682

0.621

0.711

参考文献 29

[1]	He K M， Chang H W， Sun J. Rectangling panoramic images via warping［J］. ACM Transactions on Graphics， 2013， 32（4）：79. DOI： 10.1145/2461912.2462004 .
[2]	Li D P， He K M， Sun J， et al. A geodesic-preserving method for image warping［C］//2015 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Boston， MA. IEEE， 2015：213-221.DOI： 10.1109/CVPR.2015.7298617 .
[3]	Zhang Y， Lai Y K， Zhang F L. Content-preserving image stitching with piecewise rectangular boundary constraints［J］. IEEE Transactions on Visualization and Computer Graphics， 2021， 27（7）： 3198-3212. DOI： 10.1109/TVCG.2020.2965097 .
[4]	Nie L， Lin C Y， Liao K， et al. Deep rectangling for image stitching： a learning baseline［C］//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. New Orleans， LA， USA. IEEE， 2022： 5730-5738. DOI： 10.1109/CVPR52688.2022.00565 .
[5]	Kwatra V， Schödl A， Essa I， et al. Graphcut textures： image and video synthesis using graph cuts［J］. ACM Transactions on Graphics， 2003， 22（3）： 277-286. DOI： 10.1145/882262.882264 .
[6]	Eden A， Uyttendaele M， Szeliski R. Seamless image stitching of scenes with large motions and exposure differences［C］//2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition （CVPR'06）. New York， NY， USA. IEEE， 2006： 2498-2505. DOI： 10.1109/CVPR.2006.268 .
[7]	Li N， Liao T L， Wang C. Perception-based seam cutting for image stitching［J］. Signal， Image and Video Processing， 2018， 12（5）： 967-974. DOI： 10.1007/s11760-018-1241-9 .
[8]	Gao J， Li Y， Chin T J，et al.Seam-driven image stitching［C］//Eurographics （Short Papers），2013.DOI：10.2312/CONF/EG2013/SHORT/045-048 .
[9]	Zhang F， Liu F. Parallax-tolerant image stitching［C］//2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus， OH， USA. IEEE， 2014： 3262-3269. DOI： 10.1109/CVPR.2014.423 .
[10]	Lin K M， Jiang N J， Cheong L F， et al. SEAGULL： seam-guided local alignment for parallax-tolerant image stitching［C］//European Conference on Computer Vision. Cham： Springer， 2016： 370-385.10.1007/978-3-319-46487-9_23.
[11]	Dai Q Y， Fang F M， Li J C， et al. Edge-guided composition network for image stitching［J］. Pattern Recognition， 2021， 118： 108019. DOI： 10.1016/j.patcog.2021.108019 .
[12]	Nie L， Lin C Y， Liao K， et al. Parallax-tolerant unsupervised deep image stitching［C］//2023 IEEE/CVF International Conference on Computer Vision （ICCV）. Paris， France. IEEE， 2023： 7365-7374. DOI： 10.1109/ICCV51070.2023.00680 .
[13]	Sohl-Dickstein J， Weiss E A， Maheswaranathan N， et al. Deep unsupervised learning using nonequilibrium thermodynamics［C］//Proceedings of the 32nd International Conference on International Conference on Machine Learning-Volume 37. July 6 - 11， 2015， Lille， France. ACM， 2015： 2256-2265. DOI： 10.5555/3045118.3045358 .
[14]	Ho J， Jain A， Abbeel P. Denoising diffusion probabilistic models［C］//Proceedings of the 34th International Conference on Neural Information Processing Systems. December 6 - 12， 2020，Vancouver， BC， Canada. ACM， 2020： 6840-6851. DOI： 10.5555/3495724.3496298 .
[15]	Nichol A， Dhariwal P. Improved denoising diffusion probabilistic models［EB/OL］. arXiv 2021：2102.09672. （2021-02-18）［2024-05-20］. DOI： 10.48550/arXiv.2102.09672 .
[16]	Zhu Y Z， Zhang K， Liang J Y， et al. Denoising diffusion models for plug-and-play image restoration［C］//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Vancouver， BC， Canada. IEEE， 2023： 1219-1229. DOI： 10.1109/CVPRW59228.2023.00129 .
[17]	Lugmayr A， Danelljan M， Romero A， et al. RePaint： inpainting using denoising diffusion probabilistic models［C］//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. New Orleans， LA， USA. IEEE， 2022： 11451-11461. DOI： 10.1109/CVPR52688.2022.01117 .
[18]	Chung H， Sim B， Ryu D， et al. Improving diffusion models for inverse problems using manifold constraints［C］//Proceedings of the 36th International Conference on Neural Information Processing Systems. 28 November 2022， New Orleans， LA， USA. ACM， 2022： 25683-25696. DOI： 10.5555/3600270.3602132 .
[19]	Dhariwal P， Nichol A. Diffusion models beat GANs on image synthesis［EB/OL］. arXiv 2021： 2105.05233.（2021-06-01）［2024-05-20］. .
[20]	Lowe D G. Object recognition from local scale-invariant features［C］//Proceedings of the Seventh IEEE International Conference on Computer Vision. Kerkyra， Greece. IEEE， 1999： 1150-1157. DOI： 10.1109/ICCV.1999.790410 .
[21]	Fischler M A， Bolles R C. Random sample consensus： a paradigm for model fitting with applications to image analysis and automated cartography［M］//Readings in Computer Vision. Amsterdam： Elsevier， 1987： 726-740. DOI： 10.1016/b978-0-08-051581-6.50070-2 .
[22]	Chen Y B， Liu S F， Wang X L. Learning continuous image representation with local implicit image function［C］//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville， TN， USA. IEEE， 2021： 8624-8634. DOI： 10.1109/CVPR46437.2021.00852 .
[23]	Lim B， Son S， Kim H， et al. Enhanced deep residual networks for single image super-resolution［C］//2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Honolulu， HI， USA. IEEE， 2017： 1132-1140. DOI： 10.1109/CVPRW.2017.151 .
[24]	Ke J J， Wang Q F， Wang Y L， et al. MUSIQ： multi-scale image quality transformer［C］//2021 IEEE/CVF International Conference on Computer Vision （ICCV）. Montreal， QC， Canada. IEEE， 2021： 5128-5137. DOI： 10.1109/ICCV48922.2021.00510 .
[25]	Ying Z Q， Niu H R， Gupta P， et al. From patches to pictures （PaQ-2-PiQ）： mapping the perceptual space of picture quality［C］//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Seattle， WA， USA. IEEE， 2020： 3572-3582. DOI： 10.1109/CVPR42600.2020.00363 .
[26]	Jia Q， Li Z J， Fan X， et al. Leveraging line-point consistence to preserve structures for wide parallax image stitching［C］//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville， TN， USA. IEEE， 2021： 12181-12190. DOI： 10.1109/CVPR46437.2021.01201 .
[27]	Liao T L， Li N. Single-perspective warps in natural image stitching［J］. IEEE Transactions on Image Processing， 2019，29：724-735. DOI： 10.1109/TIP.2019.2934344 .
[28]	Lin C C， Pankanti S U， Ramamurthy K N， et al. Adaptive as-natural-as-possible image stitching［C］//2015 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Boston， MA， USA. IEEE， 2015： 1155-1163. DOI： 10.1109/CVPR.2015.7298719 .
[29]	Zaragoza J， Chin T J， Brown M S， et al. As-projective-as-possible image stitching with moving DLT［C］//2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland， OR， USA. IEEE， 2013： 2339-2346. DOI： 10.1109/CVPR.2013.303 .