客觀評估一張修復後影像的品質，須同時考量三個互不等價的面向：圖像資訊保真度、人類感知品質與機器偏好，其中機器偏好反映修復結果於語義分割、物件偵測與影像檢索等下游機器視覺任務上之可用性。現有修復方法往往僅能在其中一至二個面向取得優勢，卻以犧牲其餘面向為代價：傳統回歸模型雖能維持較高的訊號保真度，卻因傾向於產生過度平滑之結果而喪失高頻細節；新興的生成式模型（如擴散模型）雖擅長合成逼真紋理，卻常伴隨隨機性幻覺與高昂的迭代運算成本，致使修復結果偏離真實影像內容並誤導機器判讀。破壞式壓縮（如 JPEG 與 WebP 等格式）為此三面向評估提供了一個兼具困難度與實務重要性的測試場景：此類演算法透過非線性量化引入區塊效應、振鈴偽影與色彩偏移等不可逆失真，使影像復原成為一高度病態的逆問題，並同時損害上述三個品質面向。因此，如何使修復方法在三者間皆達標而不偏廢，已成為當前影像復原領域之核心挑戰。
為克服上述瓶頸，本研究提出 FDIR（Flow-Guided Efficient Dual-Stage Image Restoration），一種結合生成式先驗知識與保真度精煉的雙階段修復框架，旨在不犧牲推論效率的前提下，同步優化影像之結構真實性與感知自然度。第一階段採用「壓縮率因子導引一步流匹配」，透過低秩適應 微調預訓練流匹配模型，於潛在空間中以單步推論實現壓縮感知的全域語義結構恢復。流匹配作為近年最具潛力的生成範式之一，其在影像復原領域之應用仍處於發展初期，本研究藉此系統性地探索其修復能力與適用邊界。第二階段引入「流引導條件之細節優化」，於像素空間中以確定性回歸方式進行細節修復，利用第一階段所提供之語義引導，在消除生成式幻覺的同時恢復高頻紋理，有效提升機器視覺之可辨識性。此外，本研究於效率層面同時達成兩項優勢：僅需約 3.5K 張影像即可完成訓練，且推論過程僅需兩次非迭代式前向傳遞。
經過多項實驗結果顯示，本研究是唯一能在三個品質面向上同時取得良好成績的方法：於極端壓縮（QF=1, 5）下展現最佳的保真度與具競爭力的感知品質，且其保真度優勢在整個品質區間維持穩定；於實用壓縮率（QF=20, 30）之語義分割、物件偵測與影像檢索等下游任務中，亦於大部分指標取得所比較方法中之最佳表現，其餘指標保持次佳。相較於回歸模型或生成式模型僅在單一面向領先卻於其他面向嚴重受損，FDIR 不偏好特定面向，因而在綜合三面向評估上表現最佳。本研究亦主張，僅以保真度與感知二維度評估修復方法尚不足夠——機器偏好構成獨立於傳統影像品質評估之第三評估軸，此三維評估框架本身即為本研究之方法論貢獻。本研究以破壞式壓縮此一兼具困難度與實務重要性的任務作為主要測試場景，以流匹配方法為核心達成高效且高品質的修復，並透過保真度、人類感知與機器偏好三個層面之完整評估體系提供更全面的見解，為未來影像復原與生成式模型之研究奠定堅實基礎。

A faithful evaluation of restored images requires assessing three complementary viewpoints: pixel fidelity, human perception, and machine preference. Existing methods typically optimize only one or two viewpoints at the expense of the others: regression-based approaches preserve fidelity but produce overly smooth results, while generative models synthesize realistic textures yet introduce prior-induced bias, incur high latency, and compromise fidelity, potentially misleading downstream models. This thesis focuses on images degraded by lossy compression standards such as JPEG and WebP, which are widely used for efficient storage and transmission but inevitably introduce irreversible quantization artifacts, including blocking, ringing, and color distortions. These degradations simultaneously impair all three viewpoints and make restoration a severely ill-posed inverse problem, providing a challenging and practically important task.
To reconcile these objectives, this thesis proposes FDIR (Flow-Guided Efficient Dual-Stage Image Restoration), a two-stage framework built on a structural asymmetry: perceptual plausibility benefits from generative transport toward the natural-image manifold, whereas pixel fidelity and machine preference require observation-anchored deterministic regression. The first stage, Quality-Guided One-Step Flow Matching (QO-Flow), adapts a pre-trained Rectified Flow model via Low-Rank Adaptation (LoRA) to perform compression-aware structural recovery in the latent space, achieving efficient one-step inference ($ ext{NFE}=1$) through straight-line optimal transport paths that eliminate the latency burden of iterative sampling. The second stage, Flow-Conditioned Detail Refinement (FCDR), performs deterministic pixel-space refinement by leveraging both the semantic priors from the first stage and the residual information retained in the degraded input. By anchoring the restoration to the original signal through supervised regression, FCDR effectively suppresses generative hallucinations while recovering faithful high-frequency textures and ensuring semantic consistency for downstream machine vision tasks.
Extensive experiments show FDIR achieves the highest fidelity (PSNR, SSIM) and competitive perceptual quality (LPIPS, DISTS, FID) under extreme compression ($ ext{QF}=1, 5$), with its fidelity advantage sustained across the full quality spectrum. Furthermore, at practical compression ($ ext{QF}=20, 30$), FDIR consistently outperforms competitive baselines on machine preference, our composite criterion spanning segmentation, detection, and retrieval, while requiring only approximately 3.5K training images, fewer than competing generative methods. Beyond these results, this thesis argues that machine preference is an independent third axis rather than a corollary of perceptual quality, establishing this tri-axis evaluation framework, instantiated here on lossy compression restoration, as a methodological contribution in its own right.

[1] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing, 26(7):3142–3155, July 2017.
[2] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image Restoration Using Swin Transformer, August 2021.
[3] Gregory K. Wallace. The JPEG still picture compression standard. Commun. ACM, 34(4):30–44, 1991.
[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a Deep Convolutional Network for Image Super-Resolution. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 184–199, Cham, 2014. Springer International Publishing.
[5] Yochai Blau and Tomer Michaeli. The Perception-Distortion Tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6228–6237, June 2018.
[6] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, April 2018.
[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, December 2020.
[8] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling, February 2023.
[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium, January 2018.
[10] Regev Cohen, Idan Kligvasser, Ehud Rivlin, and Daniel Freedman. Looks Too Good To Be True: An Information-Theoretic Analysis of Hallucinations in Generative Restoration Models, October 2024.
[11] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior, April 2024.
[12] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis, June 2021.
[13] Chunyi Li, Yuan Tian, Xiaoyue Ling, Zicheng Zhang, Haodong Duan, Haoning Wu, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Guo Lu, Weisi Lin, and Guangtao Zhai. Image Quality Assessment: From Human to Machine Preference, March 2025.
[14] Yiming Shi, Xiongkuo Min, Yixuan Gao, and Guangtao Zhai. Machine Vision Quality Assessment for Image Restoration. In 2025 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5, May 2025.
[15] I.-Hsiang Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, and Ming-Hsuan Yang. UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior, June 2025.
[16] Zizheng Yang, Jie Huang, Jiahao Chang, Man Zhou, Hu Yu, Jinghao Zhang, and Feng Zhao. Visual Recognition-Driven Image Restoration for Multiple Degradation with Intrinsic Semantics Recovery. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14059–14070, June 2023.
[17] Bingchen Li, Xin Li, Yiting Lu, Ruoyu Feng, Mengxi Guo, Shijie Zhao, Li Zhang, and Zhibo Chen. PromptCIR: Blind Compressed Image Restoration with Prompt Learning, April 2024.
[18] Jinpei Guo, Zheng Chen, Wenbo Li, Yong Guo, and Yulun Zhang. Compression-Aware One-Step Diffusion Model for JPEG Artifact Removal, July 2025.
[19] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. https://arxiv.org/abs/2209.03003v1, September 2022.
[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis, March 2024.
[21] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models, October 2021.
[22] A. Buades, B. Coll, and J. M. Morel. A Review of Image Denoising Algorithms, with a New One. Multiscale Modeling & Simulation, 4(2):490–530, January 2005.
[23] Maria Zontak and Michal Irani. Internal statistics of a single natural image. In CVPR 2011, pages 977–984, June 2011.
[24] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual Dense Network for Image Super-Resolution, March 2018.
[25] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, May 2017.
[26] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression Artifacts Reduction by a Deep Convolutional Network, April 2015.
[27] Jiaxi Jiang, Kai Zhang, and Radu Timofte. Towards Flexible Blind JPEG Artifacts Removal, September 2021.
[28] Woo Kyoung Han, Sunghoon Im, Jaedeok Kim, and Kyong Hwan Jin. JDEC: JPEG Decoding via Enhanced Continuous Cosine Coefficients, April 2024.
[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, April 2022.
[30] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image Super-Resolution via Iterative Refinement, June 2021.
[31] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild, April 2024.
[32] Bahjat Kawar, Jiaming Song, Stefano Ermon, and Michael Elad. JPEG Artifact Correction using Denoising Diffusion Restoration Models, November 2022.
[33] Yulin Ren, Xin Li, Bingchen Li, Xingrui Wang, Mengxi Guo, Shijie Zhao, Li Zhang, and Zhibo Chen. MoE-DiffIR: Task-customized Diffusion Priors for Universal Compressed Image Restoration, July 2024.
[34] Ségolène Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. PnP-Flow: Plug-and-Play Image Restoration with Flow Matching, September 2025.
[35] Haina Qin, Wenyang Luo, Libin Wang, Dandan Zheng, Jingdong Chen, Ming Yang, Bing Li, and Weiming Hu. Reversing Flow for Image Restoration, June 2025.
[36] Yixuan Zhu, Wenliang Zhao, Ao Li, Yansong Tang, Jie Zhou, and Jiwen Lu. FlowIE: Efficient Image Enhancement via Rectified Flow, June 2024.
[37] Guy Ohayon, Tomer Michaeli, and Michael Elad. Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration. In The Thirteenth International Conference on Learning Representations, October 2024.
[38] Zhou Wang, Alan Conrad Bovik, Hamid Rahim Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE transactions on image processing: a publication of the IEEE Signal Processing Society, 13(4):600–612, April 2004.
[39] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. Image Quality Assessment: Unifying Structure and Texture Similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2020.
[40] Wei Wang, Ruiming Guo, Yapeng Tian, and Wenming Yang. CFSNet: Toward a Controllable Feature Space for Image Restoration, August 2019.
[41] Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes, December 2022.
[42] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A. Baumann, Vincent Tao Hu, and Björn Ommer. Boosting Latent Diffusion with Flow Matching. In Computer Vision – ECCV 2024, volume 15119, pages 338–355. Springer Nature Switzerland, 2025.
[43] Dogyun Park, Taehoon Lee, Minseok Joo, and Hyunwoo J. Kim. Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation, October 2025.
[44] Shourya Verma, Mengbo Wang, Nadia Atallah Lanman, and Ananth Grama. RestoRect: Degraded Image Restoration via Latent Rectified Flow & Feature Distillation, September 2025.
[45] Jiaqi Xu, Wenbo Li, Haoze Sun, Fan Li, Zhixin Wang, Long Peng, Jingjing Ren, Haoran Yang, Xiaowei Hu, Renjing Pei, and Pheng-Ann Heng. Fast Image Super-Resolution via Consistency Rectified Flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11755–11765, 2025.
[46] Eirikur Agustsson and Radu Timofte. NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1122–1131, July 2017.
[47] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural Networks, 2(5):359–366, January 1989.
[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation, May 2015.
[49] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning, November 2017.
[50] Yuxin Wu and Kaiming He. Group Normalization, June 2018.
[51] Emily Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks, June 2015.
[52] H. Sheikh. Live image quality assessment database release 2. https://scispace.com/papers/live-image-quality-assessment-database-release-2-2a7pqf93m0, January 2005.
[53] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization, January 2019.
[54] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking Atrous Convolution for Semantic Image Segmentation, December 2017.
[55] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN, January 2018.
[56] Dillon Reis, Jordan Kupec, Jacqueline Hong, and Ahmad Daoudi. Real-Time Flying Object Detection with YOLOv8, May 2024.
[57] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011.
[58] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, January 2016.
[59] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully Convolutional One-Stage Object Detection, August 2019.
[60] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal Loss for Dense Object Detection, February 2018.
[61] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár. Microsoft COCO: Common Objects in Context, February 2015.
[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, February 2021.
[63] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning Robust Visual Features without Supervision, February 2024.
[64] Mingxing Tan and Quoc V. Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, September 2020.
[65] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, December 2015.
[66] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, August 2021.
[67] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021.
[68] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, June 2009.
[69] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a “Completely Blind” Image Quality Analyzer. IEEE Signal Processing Letters, 20(3):209–212, March 2013.
[70] Xiangpeng Tian, Xiangyu Liao, Xiao Liu, Meng Li, and Chao Ren. Degradation-Aware Feature Perturbation for All-in-One Image Restoration, May 2025.