A Test-Time Training Based Cross-Domain Image Deblurring Method

doi:10.11871/jfdc.issn.2096-742X.2026.03.010

Abstract

Abstract:

[Purpose] This study aims to address the problem of cross-domain image deblurring by proposing a novel test-time training method. [Methods] A defocus blur generation network (DBGN) is constructed by simulating the formation process of defocus blur, which is embedded at the end of the deblurring model to create an auxiliary task. During the training phase, the DBGN serves as an additional auxiliary loss to optimize the deblurring model and enhance deblurring accuracy. In the testing phase, the DBGN is utilized to perform a re-blurring task, acting as an auxiliary module to assist the primary deblurring model in updating parameters to adapt to out-of-distribution cross-domain data. [Results] The proposed method is tested on blurred images captured in inspection scenarios, validating its practical performance for cross-domain image deblurring in real-world settings. [Conclusions] Extensive experiments on multiple public defocus blur datasets and comparisons with current state-of-the-art methods demonstrate the effectiveness of the proposed approach.

Key words: image processing, image deblurring, test-time training

CHU Jingchun,YANG Guangjun,WANG Wenbin,GAO Siyuan,GAO Manda,ZHANG Sen,HE Yong. A Test-Time Training Based Cross-Domain Image Deblurring Method[J]. Frontiers of Data and Computing, 2026, 8(3): 110-121.

Figures/Tables 8

Fig.1

Table 1

Table 2

Fig. 2

Table 3

Fig.3

Table 4

Table 5

References 38

[1]	ABUOLAIM A, BROWN M S. Defocus deblurring using dual-pixel data[C]// Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. Springer International Publishing, 2020: 111-126.
[2]	汪文靖, 杨文瀚, 方玉明, 等. 恶劣场景下视觉感知与理解综述[J]. 中国图象图形学报, 2024, 29(6): 1667-1684.
[3]	RUAN L, CHEN B, LI J, et al. AIFNet: All-in-focus image restoration network using a light field-based dataset[J]. IEEE Transactions on Computational Imaging, 2021, 7: 675-688. doi: 10.1109/TCI.2021.3092891
[4]	ABUOLAIM A, AFIFI M, BROWN M S. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning[C]// Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022: 1231-1239.
[5]	郭业才, 阳刚, 毛湘南. 一种融合Transformer的多尺度结构图像去模糊方法[J]. 电光与控制, 2025, 32(3): 62-68.
[6]	BANDO Y, NISHITA T. Towards digital refocusing from a single photograph[C]// 15th Pacific Conference on Computer Graphics and Applications (PG'07), IEEE, 2007: 363-372.
[7]	SHI J, XU L, JIA J. Just noticeable defocus blur detection and estimation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 657-665.
[8]	PARK J, TAI Y W, CHO D, et al. A unified approach of multi-scale deep and hand-crafted features for defocus estimation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 1736-1745.
[9]	KARAALI A, JUNG C R. Edge-based defocus blur estimation with adaptive scale selection[J]. IEEE Transactions on Image Processing, 2017, 27(3): 1126-1137. doi: 10.1109/TIP.2017.2771563
[10]	程文涛, 任冬伟, 王旗龙. 基于循环神经网络的散焦图像去模糊算法[J]. 计算机应用研究, 2022, 39(7): 2203-2209.
[11]	D’ANDRÈS L, SALVADOR J, KOCHALE A, et al. Non-parametric blur map regression for depth of field extension[J]. IEEE Transactions on Image Processing, 2016, 25(4): 1660-1673. doi: 10.1109/TIP.2016.2526907 pmid: 26886992
[12]	CHO S, LEE S. Convergence analysis of MAP based blur kernel estimation[C]// Proceedings of the IEEE International Conference on Computer Vision, 2017: 4808-4816.
[13]	LEE J, LEE S, CHO S, et al. Deep defocus map estimation using domain adaptation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 12222-12230.
[14]	FISH D A, BRINICOMBE A M, PIKE E R, et al. Blind deconvolution by means of the Richardson-Lucy algorithm[J]. Journal of the Optical Society of America A, 1995, 12(1): 58-65. doi: 10.1364/JOSAA.12.000058
[15]	KRISHNAN D, FERGUS R. Fast image deconvolution using hyper-Laplacian priors[J]. Advances in Neural Information Processing Systems, 2009, 22: 1-9.
[16]	胡张颖, 周全, 陈明举, 等. 图像去模糊研究综述[J]. 中国图象图形学报, 2024, 29(4): 841-861.
[17]	刘利平, 孙建, 高世妍. 单图像盲去模糊方法概述[J]. 计算机科学与探索, 2022, 16(3): 552-564. doi: 10.3778/j.issn.1673-9418.2106100
[18]	SON H, LEE J, CHO S, et al. Single image defocus deblurring using kernel-sharing parallel atrous convolutions[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 2642-2650.
[19]	LEE J, SON H, RIM J, et al. Iterative filter adaptive network for single image defocus deblurring[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 2034-2042.
[20]	QUAN Y, WU Z, JI H. Gaussian kernel mixture network for single image defocus deblurring[J]. Advances in Neural Information Processing Systems, 2021, 34: 20812-20824.
[21]	RUAN L, CHEN B, LI J, et al. Learning to deblur using light field generated and real defocus images[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 16304-16313.
[22]	CHEN H, GU J, GALLO O, et al. Reblur2deblur: Deblurring videos via self-supervised learning[C]// 2018 IEEE International Conference on Computational Photography (ICCP), IEEE, 2018: 1-9.
[23]	MA H, LIU S, LIAO Q, et al. Defocus image deblurring network with defocus map estimation as auxiliary task[J]. IEEE Transactions on Image Processing, 2021, 31: 216-226. doi: 10.1109/TIP.2021.3127850
[24]	LI Y, REN D, SHU X, et al. Learning single image defocus deblurring with misaligned training pairs[C]// Proceedings of the AAAI Conference on Artificial Intelligence, 2023, 37(2): 1495-1503.
[25]	朱金秀, 徐传蕾, 朱京京, 等. 基于注意力机制和提示学习的图像去模糊网络[J]. 计算机测量与控制, 2025, 33(9): 310-317+325.
[26]	陈康, 林建涵, 刘元杰. 图像去模糊算法研究综述[J]. 计算机科学, 2025, 52(11): 98-112.
[27]	SUN Y, WANG X, LIU Z, et al. Test-time training with self-supervision for generalization under distribution shifts[C]// International Conference on Machine Learning, PMLR, 2020: 9229-9248.
[28]	LIU H, WU Z, LI L, et al. Towards multi-domain single image dehazing via test-time training[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 5831-5840.
[29]	GUNAWAN A, NUGROHO M A, PARK S J. Test-time adaptation for real image denoising via meta-transfer learning[DB/OL]. arXiv preprint, 2022. https://arxiv.org/abs/2207.02066.
[30]	CHI Z, WANG Y, YU Y, et al. Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 9137-9146.
[31]	SOH J W, CHO S, CHO N I. Meta-transfer learning for zero-shot super-resolution[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 3516-3525.
[32]	RONNEBERGER O, FISCHER P, BROX T. U-net: Convolutional networks for biomedical image segmentation[C]// Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, Springer International Publishing, 2015: 234-241.
[33]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
[34]	WANG Z, BOUIK 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. doi: 10.1109/tip.2003.819861 pmid: 15376593
[35]	ZHANG R, ISOLA P, EFROS A A, et al. The unreasonable effectiveness of deep features as a perceptual metric[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 586-595.
[36]	LIU L, JIANG H, HE P, et al. On the variance of the adaptive learning rate and beyond[DB/OL]. arXiv preprint, 2019. https://arxiv.org/abs/1908.03265.
[37]	HARDT M, RECHT B, SINGER Y. Train faster, generalize better: Stability of stochastic gradient descent[C]// International Conference on Machine Learning, PMLR, 2016: 1225-1234.
[38]	ZAMIR S W, ARORA A, KHAN S, et al. Restormer: Efficient transformer for high-resolution image restoration[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 5728-5739.

Methods	DPDD
Methods	PSNR（↑）	SSIM（↑）	LPIPS（↓）
Blurry Input	23.89 dB	0.725	0.471
KPAC^[18]	25.22 dB	0.774	0.226
IFAN^[19]	25.37 dB	0.789	0.217
GKMNet^[20]	25.36 dB	0.774	0.276
DRBNet^[21]	25.47 dB	0.787	0.246
Ours（F_MDN-L+F_DBGN）	26.03 dB	0.812	0.154

Method	RealDoF			LF-DoF			DED			RTF			SDD
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Blurry Input	22.33	0.633	0.524	25.87	0.779	0.316	28.58	0.884	0.164	24.18	0.739	0.364	23.09	0.679	0.417
KPAC^[18]	23.98	0.716	0.336	25.40	0.790	0.225	28.16	0.890	0.114	24.57	0.815	0.233	22.18	0.672	0.357
IFAN^[19]	24.71	0.749	0.306	26.11	0.817	0.220	27.85	0.888	0.111	24.92	0.821	0.215	22.53	0.680	0.377
GKMNet^[20]	24.30	0.716	0.399	25.96	0.802	0.271	27.93	0.883	0.144	25.10	0.826	0.274	22.98	0.685	0.406
Restormer^[38]	25.08	0.769	0.289	26.44	0.823	0.207	26.95	0.884	0.113	24.21	0.822	0.204	22.59	0.685	0.497
Ours	24.97	0.768	0.262	26.70	0.824	0.191	28.16	0.891	0.101	25.49	0.847	0.171	22.84	0.680	0.356

Model	Params (Mb)	Speed（fps）
KPAC^[18]	1.58	17.8
IFAN^[19]	10.48	9.8
GKMNet^[20]	1.41	23.2
DRBNet^[21]	10.69	8.7
Restormer^[38]	26.10	4.3
Ours	19.12	7.6

Variants	L_deblur	L_reblur	Our TTT	MA-TTT in [30]	DPDD	RealDoF	LF-DoF	DED	RTF	SDD
V₁	$\surd $				25.69	24.43	26.25	27.91	25.06	22.76
V₂	$\surd $	$\surd $			25.72	24.50	26.33	28.12	25.30	22.83
V₃	$\surd $			$\surd $	25.64	24.32	26.15	27.98	24.99	22.79
V₄	$\surd $	$\surd $	$\surd $		25.73	24.48	26.44	28.30	25.44	22.87

	DBGN w/o DM	DBGN w/o ResBlocks	DBGN
DPDD	25.73	25.73	25.74
RealDoF	24.52	24.28	27.54
LF-DoF	26.42	26.40	26.44
DED	28.18	28.22	28.30
RTF	25.02	25.10	25.44
SDD	22.81	22.87	22.87