請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83247
標題: | 無特定假設之自監督式影像去噪 Towards Assumption-Free and Self-Supervised Image Denoising |
其他標題: | Towards Assumption-Free and Self-Supervised Image Denoising |
作者: | 陳宇軒 Yu-Hsuan Chen |
指導教授: | 王鈺強 Yu-Chiang Wang |
關鍵字: | 影像去噪,深度學習,自監督式學習,對抗式機器學習, image denoising,deep learning,self-supervised learning,adversarial learning, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 影像去噪是電腦視覺領域中一項基本的底層視覺任務,目的為移除影像中不必要的雜訊,從噪聲影像中還原出無噪聲的信號。目前基於深度學習的方法已展現出良好的去噪結果,然而由於其訓練資料的要求與預定義的假設,這些方法對真實世界資料的適用性可能受到限制。為了緩解這些限制,我們提出了一個針對真實世界影像去噪的自監督式學習框架,只需要使用單一未配對的噪聲影像就能進行模型訓練,同時不需要對噪聲分布或圖像分解作出特定的假設。在此一自監督式學習框架中,我們引入了兩個學習階段:自監督式的信號恢復,以及信號感知的噪聲生成。前者利用預測之無噪聲和噪聲增強樣本,建立多個循環式影像重建的學習目標來訓練去噪器;後者則學習捕捉與模擬訓練資料中的噪聲分布,以產生信號感知的偽配對影像,來作為額外的訓練監督。我們針對真實世界噪聲影像以及噪聲合成資料集進行了廣泛的實驗,相較於近期最先進的無監督和自監督去噪方法,我們的學習框架在質化與量化結果上皆能有更好的表現,驗證了此一自監督式學習框架的有效性和適用性。 Image denoising is a fundamental low-level vision task that aims to restore clean signals from noisy images. While recent deep learning-based approaches have shown promising results, their applicability to real-world data might be limited due to their training data requirements or predefined assumptions. To alleviate such limitations, we propose Noisy-to-Clean-to-Noisy (N2C2N), a self-supervised learning scheme for real-world denoising, which learns from single noisy images without making particular assumptions on noise distribution or image decomposition. We introduce two self-supervised learning stages in N2C2N: self-supervised signal recovery and signal-aware noise generation. The former trains the denoiser by encouraging various cyclic image reconstruction consistencies with the predicted noise-free and noise-augmented samples, and the latter exploits underlying noise distribution and produces pseudo noisy-denoised image pairs as additional training supervision. Extensive experiments on both real-world noisy images and synthetic datasets verify the effectiveness and applicability of our proposed learning framework, which is shown to perform favorably against state-of-the-art unsupervised and self-supervised denoising methods. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83247 |
DOI: | 10.6342/NTU202300063 |
全文授權: | 未授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0745230110252088.pdf 目前未授權公開取用 | 34.06 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。