基於色彩導引的控制網路與擴散模型之盲影像去模糊方法;Blind Image Deblurring via Chromatic-Guided ControlNet and Latent Diffusion Models

NCU Institutional Repository > 資訊電機學院 > 軟體工程研究所 > 博碩士論文 > Item 987654321/98020

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98020

题名:	基於色彩導引的控制網路與擴散模型之盲影像去模糊方法;Blind Image Deblurring via Chromatic-Guided ControlNet and Latent Diffusion Models
作者:	柳瀚揚;Liou, Han-Yang
贡献者:	軟體工程研究所
关键词:	影像去模糊;擴散模型;生成導引;Image Deblurring;Diffusion Models;Generative Guidance
日期:	2025-06-26
上传时间:	2025-10-17 12:15:45 (UTC+8)
出版者:	國立中央大學
摘要:	影像去模糊在電腦視覺領域中仍然是一項極具挑戰性的任務，尤其在動態運動與空間變化複雜的真實場景中更為困難。近年大型擴散模型於各類影像生成任務中展現出卓越的性能，其於影像去模糊領域的應用亦逐漸受到重視，逐漸成為高品質影像重建的主流技術之一。然而，直接使用大型擴散模型於去模糊任務，不僅面臨高昂的計算資源消耗，亦因缺乏針對性條件生成設計而導致在重新繪製清晰影像時容易生成在原始影像中沒有的物件，或是重繪出一幅無關的影像。針對上述問題，本研究結合潛在擴散模型以及控制網路（ControlNet）之架構提出一個有效的生成引導方法，在繼承潛在擴散模型出色的泛化能力，與控制網路運算資源效率的同時進一步強化了模型對於生成方向的控制能力。同時，本研究設計之色彩資訊導引（Chromatic tile）模組不僅提供局部對比與色彩還原的輔助訊息，更能有效引導擴散過程中細節的重建與結構的一致性，克服傳統結構提示易忽略色彩資訊的侷限，有效緩解生成影像失去彩度的問題。本研究於真實世界模糊資料集與合成模糊資料集中進行廣泛測試，涵蓋室內與室外多樣場景。實驗結果顯示，所提方法在多個感知指標上與現有先進方法旗鼓相當，甚至在部分指標上取得更優異的去模糊效果。同時本方法能夠在NVIDIA RTX 3090 GPU上進行訓練，且推論階段僅需10 GB之顯示卡記憶體，展現良好的實用性與資源效率。 ;Image deblurring remains a highly challenging task in the field of computer vision, especially in real-world scenes with complex motion and spatial variations. Recently, diffusion models have demonstrated outstanding performance across various image generation tasks, and their application in image deblurring has increasingly gained attention, becoming one of the mainstream techniques for high-quality image reconstruction. However, directly applying diffusion models for deblurring tasks not only incurs significant computational resource demands but also risks generating irrelevant or nonexistent objects in reconstructed clear images due to a lack of targeted conditional generation design. To address these issues, we propose an effective generative guidance method by combining the latent diffusion model with ControlNet architecture. The method inherits the superior generalization capabilities of latent diffusion models and the computational efficiency of ControlNet while significantly enhancing the model′s controllability over the generation process. Furthermore, this research introduces a specially designed Chromatic Guidance module, which provides auxiliary information for local contrast and color restoration, effectively guiding detail reconstruction and structural consistency during the diffusion process. This approach overcomes the limitation of traditional structural guidance methods, which often neglect color information, resulting in effectively alleviating the issue of generated images suffering from desaturation. We have conducted extensive experiments on both real-world and synthetic blur datasets, covering diverse indoor and outdoor scenes. Experimental results indicate that the proposed method achieves comparable performance to existing well-known techniques across multiple metrics, even surpassing them on certain metrics. Moreover, our approach demonstrates excellent practicality and resource efficiency, capable of training on single NVIDIA RTX 3090 GPU and requiring only 10 GB of VRAM to inference.
显示于类别:	[軟體工程研究所 ] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	10	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....