可適性盲式影像去噪與去模糊之深度網路與潛空間球面插值方法

Abstract

在同時存在模糊與雜訊的情況下進行盲影像復原，特別是在低訊噪比條件下，仍是電腦視覺領域中的一大挑戰。這是因為在低訊噪比下，影像中的細節容易被雜訊所干擾。儘管目前許多影像去模糊方法在無雜訊時表現良好，但當影像受到雜訊干擾時，這些方法的效能會顯著下降，限制了它們在夜間攝影、監控影像或醫學影像等真實場景中的應用。 本論文提出一個可適性的影像復原框架，融合了深度神經網路的去雜訊與去模糊功能，並且無需針對不同雜訊程度進行重複訓練。我們發現，如果以「先去雜訊後去模糊」的兩階段流程進行處理，在低訊噪比時其結果並不理想。為了解決上述問題，我們訓練了兩個分別針對高雜與低雜輸入優化的去模糊模型。針對中等雜訊程度的影像，透過兩個模型在潛空間中進行球面內插，實現可適性的影像復原。 此外，本研究亦包含雜訊與模糊程度之估計模型，能判斷個別輸入影像的情況，使模型動態選擇適當的內插比例。為了進行可控且一致的測試，我們於GoPro 資料集上加入不同程度的加性高斯白雜訊(AWGN)。 從量化評估指標如PSNR 與 SSIM 的結果可看出，本方法在各種雜訊條件下皆能展現具競爭力的效能，不僅優於兩階段之方法，也優於現今其他先進的研究。儘管本研究採用合成雜訊，其展現出良好的可解釋性以及對不同劣化條件的可適能力，為混合型影像劣化情境下的盲復原任務提供一套實用且高效的解法。

Blind image restoration in the presence of both blur and noise remains a challenging problem in computer vision, particularly under low signal-to-noise ratio (SNR) conditions where fine image structures are easily overwhelmed by noise. While many existing deblurring methods perform well on clean inputs, their effectiveness drops significantly when the image is degraded by noise, limiting their applicability in real- world scenarios such as night photography, surveillance, or medical imaging. This thesis proposes an adaptive image restoration framework that integrates denoising and deblurring using deep networks, without requiring exhaustive retraining across noise levels. We begin with a two-stage pipeline: denoising followed by deblurring, as a baseline and demonstrate its limitations in preserving fine image details under low SNR. To address these limitations, we fine-tune two separate deblurring models: one optimized for high-noise inputs and another for low-noise inputs. For intermediate noise levels, we employ spherical interpolation in the latent space between the two models, enabling a continuous and adaptive restoration process without additional training overhead. Our framework also includes noise and blur estimation modules to guide the interpolation dynamically, allowing the system to adjust to unknown and varying degradation levels. For controlled evaluation, we synthetically add Additive White Gaussian Noise (AWGN) to the GoPro dataset, ensuring consistent testing across different noise scenarios. Quantitative results based on Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) show that our method delivers competitive performance compared to baselines and state-of-the-art methods. Despite using synthetic noise, the proposed framework demonstrates a scalable, interpretable, and noise-aware strategy for adaptive blind image restoration under mixed and challenging degradation conditions.