复数域幅相协同的轻量频域超分辨率网络

doi:10.3785/j.issn.1008-973X.2026.09.013

浙江大学学报(工学版)

2026, Vol. 60

Issue (9): 1962-1971 DOI: 10.3785/j.issn.1008-973X.2026.09.013

计算机技术、自动控制技术

复数域幅相协同的轻量频域超分辨率网络

崔鹏(

),高闯,孟庆涵

哈尔滨理工大学图像处理实验室，黑龙江哈尔滨 150001

Lightweight frequency-domain super-resolution network based on amplitude-phase collaboration in complex domain

Peng CUI(

),Chuang GAO,Qinghan MENG

Image Processing Laboratory, Harbin University of Science and Technology, Harbin 150001, China

全文: PDF(3121 KB) HTML

摘要：

针对频域超分辨率中实值化编码破坏复数结构完整性，导致相位失真与几何保真度下降的问题，提出幅相协同建模原则. 在复数域直接建模幅度与相位的耦合关系，显式满足实值图像傅里叶变换的共轭对称性约束. 构建轻量级频域超分辨率网络（APNet）,设计频域幅相注意力机制（FAPA），在复数域同步学习幅度缩放与相位校准，避免实部、虚部分离引起的相位漂移. 提出多尺度融合卷积模块（HSKB），增强多尺度及局部与全局上下文融合能力. 引入频域动态归一化（FDN），依据局部信噪比自适应调节复数空间归一化强度. 实验表明，在比例因子为4的Urban100和Manga109中，相较于最新方法RepRFN在参数量减少20.9%的情况下，PSNR分别提高了0.42、0.3 dB.

关键词： 单幅图像超分辨率; 轻量级; 频域; 相位; 幅度

Abstract:

An amplitude-phase collaborative modeling principle was proposed in order to address the problem that real-valued encoding destroys the integrity of complex structure in frequency-domain super-resolution, leading to phase distortion and degraded geometric fidelity. The coupling relationship between amplitude and phase was directly modeled in the complex domain, with explicit satisfaction of the conjugate symmetry constraint of Fourier transform for real-valued images. A lightweight frequency-domain super-resolution network, namely amplitude-phase network (APNet), was constructed. A frequency-domain amplitude-phase attention (FAPA) mechanism was designed to synchronously learn amplitude scaling and phase calibration in the complex domain, thereby avoiding phase drift caused by the separation of real and imaginary parts. A hierarchical scale-fused kernel block (HSKB) was proposed to enhance the capability of multi-scale and local-global context fusion. Frequency-domain dynamic normalization (FDN) was introduced to adaptively adjust the normalization intensity of the complex space according to the local signal-to-noise ratio. Experiments were conducted on Urban100 and Manga109 datasets with a scale factor of 4. The parameter count of APNet was reduced by 20.9% compared with the state-of-the-art method RepRFN, while the peak signal-to-noise ratio (PSNR) was improved by 0.42 dB and 0.3 dB on the two datasets, respectively.

Key words: single image super-resolution lightweight frequency-domain phase amplitude

收稿日期: 2025-10-06 出版日期: 2026-07-20

CLC:

TP 391

作者简介: 崔鹏（1971—），男，副教授，从事图形图像的处理与研究. oricd.org/0009-0002-7823-8919. E-mail：cuipeng83@163.com

	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
	崔鹏
	高闯
	孟庆涵

引用本文:

崔鹏,高闯,孟庆涵. 复数域幅相协同的轻量频域超分辨率网络[J]. 浙江大学学报(工学版), 2026, 60(9): 1962-1971.

Peng CUI,Chuang GAO,Qinghan MENG. Lightweight frequency-domain super-resolution network based on amplitude-phase collaboration in complex domain. Journal of ZheJiang University (Engineering Science), 2026, 60(9): 1962-1971.

链接本文:

https://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2026.09.013 或 https://www.zjujournals.com/eng/CN/Y2026/V60/I9/1962

图 1 APNet 网络的整体架构

图 2 多尺度融合卷积模块

图 3 提出的频域幅相注意力模块

表 1 频域幅相注意力模块的消融实验

表 2 多尺度融合卷积的消融实验

表 3 归一化模块的消融实验

表 4 缩放因子为4 时的各项指标对比

表 5 缩放因子为 4 时的L1误差和余弦误差对比

图 4 不同方法的SR模型的结果比较

图 5 Urban100 数据集下的性能与复杂度对比

方法	尺度因子	N_p/10³	FLOPs/10⁹	Set5		Set14		B100		Urban100		Manga109
方法	尺度因子	N_p/10³	FLOPs/10⁹	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM
DBNet^[34]	2	416	84.15	38.00	0.9602	33.43	0.9173	32.22	0.9007	32.07	0.9273	39.00	0.9775
LKDN^[21]	2	291	66.60	38.06	0.9609	34.00	0.9207	32.28	0.9011	32.92	0.9350	39.12	0.9779
RepRFN^[35]	2	386	85.12	38.07	0.9612	33.63	0.9184	32.22	0.9009	32.10	0.9274	39.00	0.9774
SAFMN^[33]	2	228	52.00	38.00	0.9611	33.54	0.9177	32.16	0.8995	31.84	0.9256	38.71	0.9771
OSFFNet^[36]	2	516	83.20	38.10	0.9601	33.72	0.9190	32.29	0.9012	32.67	0.9331	39.09	0.9780
EARFA^[37]	2	1026	229.0	38.04	0.9614	33.98	0.9212	32.32	0.9012	32.88	0.9359	39.16	0.9782
DiMoSR^[38]	2	338	76.00	38.06	0.9607	33.74	0.9194	32.24	0.9006	32.30	0.9295	38.05	0.9782
DSCLoRA^[39]	2	398	77.00	38.02	0.9607	33.72	0.9212	32.26	0.8992	32.81	0.9290	38.01	0.9678
IBMDB^[40]	2	622	92.83	38.06	0.9673	33.66	0.9174	32.16	0.8992	32.17	0.9277	38.81	0.9772
本文方法	2	306	21.00	38.15	0.9612	34.04	0.9218	32.35	0.9015	32.95	0.9353	39.16	0.9783
DBNet^[34]	3	826	70.20	34.46	0.9279	30.42	0.8427	29.18	0.8063	28.51	0.8571	33.99	0.9466
LKDN^[21]	3	311	31.40	34.54	0.9285	30.52	0.8455	29.21	0.8078	28.50	0.8601	34.08	0.9475
RepRFN^[35]	3	392	38.40	34.45	0.9280	30.39	0.8430	29.13	0.8068	28.06	0.8494	33.76	0.9451
SAFMN^[33]	3	233	23.00	34.34	0.9267	30.33	0.8418	29.08	0.8048	27.95	0.8474	33.52	0.9437
OSFFNet^[36]	3	524	37.80	34.58	0.9287	30.48	0.8450	29.21	0.8078	28.50	0.8601	34.08	0.9475
EARFA^[37]	3	1034	102.40	34.53	0.9277	30.51	0.8451	29.29	0.8083	28.87	0.8632	34.18	0.9464
DiMoSR^[38]	3	—	45.00	34.46	0.9279	30.44	0.8429	29.20	0.8073	28.51	0.8574	33.89	0.9469
DSCLoRA^[39]	3	416	25.83	34.51	0.9281	30.43	0.8444	29.06	0.8066	28.33	0.8563	34.04	0.9436
IBMDB^[40]	3	991	67.54	34.45	0.9276	30.44	0.8445	29.15	0.8065	28.37	0.8563	34.08	0.9439
本文方法	3	310	21.00	34.59	0.9286	30.54	0.8455	29.28	0.8081	28.87	0.8631	34.18	0.9476
DBNet^[34]	4	832	51.80	32.29	0.8961	28.71	0.7834	27.66	0.7377	26.34	0.7909	30.83	0.9111
LKDN^[21]	4	322	18.30	32.39	0.8979	28.79	0.7859	27.69	0.7402	26.40	0.7965	30.43	0.9140
RepRFN^[35]	4	402	22.10	32.28	0.8969	28.68	0.7836	27.65	0.7389	26.18	0.7858	30.79	0.9102
SAFMN^[33]	4	240	14.00	32.18	0.8948	28.60	0.7813	27.58	0.7359	25.97	0.7809	30.89	0.9063
OSFFNet^[36]	4	537	22.00	32.39	0.8976	28.75	0.7852	27.66	0.7393	26.36	0.7950	30.84	0.9125
EARFA^[37]	4	1209	11.61	32.38	0.8975	28.76	0.7859	27.65	0.7431	26.60	0.8004	31.07	0.9147
DiMoSR^[38]	4	349	20.00	32.31	0.8962	28.74	0.7844	27.65	0.7384	26.25	0.7889	30.89	0.9061
DSCLoRA^[39]	4	426	27.87	32.19	0.8958	28.63	0.7830	27.61	0.7370	26.13	0.7864	30.60	0.9095
IBMDB^[40]	4	917	43.56	32.23	0.8954	28.61	0.7812	27.55	0.7356	26.10	0.7856	30.71	0.9073
本文方法	4	318	21.00	32.44	0.8979	28.83	0.7864	27.69	0.7428	26.60	0.8004	31.09	0.9146

表 6 不同尺度因子的超分辨率重建模型在各测试集上的峰值信噪比与结构相似性指标对比结果

图 6 标准测试集下 4 倍放大倍数下的视觉效果比较

1	WANG Y, SHAO Z, LU T, et al Remote sensing image super-resolution via multiscale enhancement network[J]. IEEE Geoscience and Remote Sensing Letters, 2023, 20: 5000905 doi: 10.1109/lgrs.2024.3481645
2	李浦瑛. 基于深度学习的视频超分辨重建技术研究 [D]. 哈尔滨: 黑龙江大学, 2025. LI Puying. Research on video super-resolution reconstruction technology based on deep learning [D]. Harbin: Heilongjiang University, 2025.
3	刘杰, 吴优, 田佳禾, 等改进Transformer的肺部CT图像超分辨率重建[J]. 浙江大学学报: 工学版, 2025, 59 (7): 1434- 1442 LIU Jie, WU You, TIAN Jiahe, et al Based on improved Transformer for super-resolution reconstruction of lung CT images[J]. Journal of Zhejiang University: Engineering Science, 2025, 59 (7): 1434- 1442
4	DONG C, LOY C C, HE K, et al Image super-resolution using deep convolutional networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38 (2): 295- 307 doi: 10.1109/TPAMI.2015.2439281
5	KIM J, LEE J K, LEE K M. Accurate image super-resolution using very deep convolutional networks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 1646–1654.
6	LIN Z, GARG P, BANERJEE A, et al. Revisiting RCAN: improved training for image super-resolution [EB/OL]. [2022-01-28]. https://arxiv.org/abs/2201.11279.
7	XU Z, XU J, SONG B, et al. An improved compression algorithm based on IDN model of image super-resolution reconstruction [C]//Proceedings of the International Conference on Neural Networks, Information, and Communication Engineering. Qingdao: SPIE, 2022: 4.
8	LIU J, TANG J, WU G. Residual feature distillation network for lightweight image super-resolution [C]//Proceedings of the European Conference on Computer Vision Workshops. Glasgow: Springer, 2020: 41–55.
9	LIANG J, CAO J, SUN G, et al. SwinIR: image restoration using swin transformer [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. Montreal: IEEE, 2021: 1833–1844.
10	WANG X, GIRSHICK R, GUPTA A, et al. Non-local neural networks [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 7794–7803.
11	HUANG Z, ZHANG Z, LAN C, et al. Adaptive frequency filters as efficient global token mixers [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris: IEEE, 2023: 6026–6036.
12	RAO Y, ZHAO W, ZHU Z, et al GFNet: global filter networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45 (9): 10960- 10973
13	FAN D, LI H, ZHANG K, et al. FDANet: a frequency-domain attention network for super-resolution image quality assessment [EB/OL]. [2023-03-15]. https://arxiv.org/abs/2303.08934.
14	QIAO C, LI D, GUO Y, et al Evaluation and development of deep neural networks for image super-resolution in optical microscopy[J]. Nature Methods, 2021, 18 (2): 194- 202 doi: 10.1038/s41592-020-01048-5
15	ZHANG D, HUANG F, LIU S, et al. SwinFIR: revisiting the SwinIR with fast Fourier convolution and improved training for image super-resolution [EB/OL]. [2022-08-24]. https://arxiv.org/abs/2208.11247.
16	CHI L, JIANG B R, MU Y D. Fast Fourier convolution [C]//Advances in Neural Information Processing Systems 33. [S. l.]: Curran Associates, Inc., 2020: 4479–4488.
17	JENEFA A, KURIAKOSE B M, NAVEEN E V, et al EDSR: empowering super-resolution algorithms with high-quality DIV2K images[J]. Intelligent Decision Technologies, 2023, 17 (4): 1249- 1263 doi: 10.3233/IDT-230218
18	XIE X, ZHOU P, LI H, et al Adan: adaptive nesterov momentum algorithm for faster optimizing deep models[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46 (12): 9508- 9520 doi: 10.1109/TPAMI.2024.3423382
19	LOSHCHILOV I, HUTTER F. SGDR: stochastic gradient descent with warm restarts [EB/OL]. [2025-09-20]. https://arxiv.org/abs/1608.03983.
20	LI Z, LIU Y, CHEN X, et al. Blueprint separable residual network for efficient image super-resolution [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New Orleans: IEEE, 2022: 832–842.
21	XIE C, ZHANG X, LI L, et al. Large kernel distillation network for efficient single image super-resolution [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Vancouver: IEEE, 2023: 1283–1292.
22	HUI Z, GAO X, YANG Y, et al. Lightweight image super-resolution with information multi-distillation network [C]//Proceedings of the 27th ACM International Conference on Multimedia. Nice France: ACM, 2019: 2024–2032.
23	CHEN J, DUANMU C, LONG H. Large kernel frequency-enhanced network for efficient single image super-resolution [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Seattle: IEEE, 2024: 6317–6326.
24	RUSSO A E, KIRBY W M, RUDINGER K M, et al Consistency testing for robust phase estimation[J]. Physical Review A, 2021, 103 (4): 042609 doi: 10.1103/PhysRevA.103.042609
25	ZHAO H, KONG X, HE J, et al. Efficient image super-resolution using pixel attention [C]//Proceedings of the 16th European Conference on Computer Vision Workshops. Glasgow: Springer, 2020: 56–72.
26	CHEN X, WANG X, ZHANG W, et al. HAT: hybrid attention transformer for image restoration [EB/OL]. [2023-09-11]. https://arxiv.org/abs/2309.05239.
27	GU J, DONG C. Interpreting super-resolution networks with local attribution maps [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville: IEEE, 2021: 9195–9204.
28	TAI Y, YANG J, LIU X. Image super-resolution via deep recursive residual network [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 2790–2798.
29	TAI Y, YANG J, LIU X, et al. MemNet: a persistent memory network for image restoration [C]//Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 4549–4557.
30	LAI W S, HUANG J B, AHUJA N, et al. Deep Laplacian pyramid networks for fast and accurate super-resolution [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 5835–5843.
31	BEHJATI P, RODRIGUEZ P, FERNÁNDEZ C, et al Single image super-resolution based on directional variance attention network[J]. Pattern Recognition, 2023, 133: 108997 doi: 10.1016/j.patcog.2022.108997
32	DONG C, LOY C C, TANG X. Accelerating the super-resolution convolutional neural network [C]//Proceedings of the 14th European Conference on Computer Vision. Amsterdam: Springer, 2016: 391–407.
33	SUN L, DONG J, TANG J, et al. Spatially-adaptive feature modulation for efficient image super-resolution [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris: IEEE, 2024: 13144–13153.
34	GU R, SHEN Z, CHEN J. Single image super-resolution reconstruction technique with lightweight dual-branch network [C]//Proceedings of the 8th International Conference on Image, Vision and Computing. Dalian: IEEE, 2023: 565–571.
35	DENG W, YUAN H, DENG L, et al. Reparameterized residual feature network for lightweight image super-resolution [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Vancouver: IEEE, 2023: 1712–1721.
36	WANG Y, ZHANG T OSFFNet: omni-stage feature fusion network for lightweight image super-resolution[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38 (6): 5660- 5668 doi: 10.1609/aaai.v38i6.28377
37	ZHAO X, LI L, XIE C, et al. Efficient single image super-resolution with entropy attention and receptive field augmentation [C]//Proceedings of the 32nd ACM International Conference on Multimedia. Melbourne: ACM, 2024: 1302–1310.
38	YILMAZ M A, BILICAN A, TEKALP A M. DiMOSR: feature modulation via multi-branch dilated convolutions for efficient image super-resolution [EB/OL]. [2025-05-27]. https://arxiv.org/abs/2505.21262.
39	CHAI X, ZHANG Y, ZHANG Y, et al. Distillation-supervised convolutional low-rank adaptation for efficient image super-resolution [EB/OL]. [2025-04-15]. https://arxiv.org/abs/2504.11271.

[1]	杨燕,宋鑫钰. 基于动态频域调制的交互式图像去雾网络[J]. 浙江大学学报(工学版), 2026, 60(6): 1221-1230.
[2]	周华平,邓彬,孙克雷,张咏琪,吴涛,吴劲. 井下轨道区域高精度实时分割网络[J]. 浙江大学学报(工学版), 2026, 60(2): 332-340.
[3]	邱文昊,卢旸,成亮,姚陈,翟彦蓉,毕美华. 新型多幅度编码在再调制无源光网络中的应用[J]. 浙江大学学报(工学版), 2025, 59(2): 433-441.
[4]	杨钢,潘越,王照卓,徐越,李宝仁. 电液伺服系统线性自抗扰控制参数频域整定方法[J]. 浙江大学学报(工学版), 2025, 59(10): 2034-2044.
[5]	王光明,柏正尧,宋帅,徐月娥. 阿尔茨海默病辅助诊断的多模态数据融合轻量级网络[J]. 浙江大学学报(工学版), 2025, 59(1): 39-48.
[6]	张松,傅海鹏. 2.4 GHz GaAs HBT高线性度功率放大器设计[J]. 浙江大学学报(工学版), 2024, 58(7): 1524-1532.
[7]	罗伟,颜作涛,关佳浩,韩建. 基于改进SegFormer的太阳能电池缺陷分割模型[J]. 浙江大学学报(工学版), 2024, 58(12): 2459-2468.
[8]	冯志成,杨杰,陈智超. 基于轻量级Transformer的城市路网提取方法[J]. 浙江大学学报(工学版), 2024, 58(1): 40-49.
[9]	郭浩然,郭继昌,汪昱东. 面向水下场景的轻量级图像语义分割网络[J]. 浙江大学学报(工学版), 2023, 57(7): 1278-1286.
[10]	贺俊,张雅声,尹灿斌. 基于深度学习的星载SAR工作模式鉴别[J]. 浙江大学学报(工学版), 2022, 56(8): 1676-1684.
[11]	徐砚天,黄晓敏,李浩明,王志宇,郁发新. 多芯片小数分频锁相环输出信号相位同步设计[J]. 浙江大学学报(工学版), 2021, 55(9): 1788-1794.
[12]	李洋,齐蓉,代明光. 补偿因子可调逆变器电压外环线性自抗扰控制[J]. 浙江大学学报(工学版), 2021, 55(7): 1279-1288.
[13]	高向宇,李杨龙. 利用AR模型解决全相位技术数据长度限制[J]. 浙江大学学报(工学版), 2021, 55(6): 1108-1117.
[14]	高新智,刘作军,张燕,陈玲玲. 基于GWO-SVM的下肢假肢穿戴者骑行相位识别[J]. 浙江大学学报(工学版), 2021, 55(4): 648-657.
[15]	程鸿,胡佳杰,刘勇,叶远青. 强度传输方程和神经网络融合的三维重构算法[J]. 浙江大学学报(工学版), 2021, 55(4): 658-664.

Viewed

Full text

Abstract

Cited

Shared

Discussed