閃光燈引導影像去噪旨在利用閃光燈影像所提供的結構資訊，從含噪的環境光 影像中還原出清晰的結果。然而，由於閃光燈影像與環境光影像在光度外觀上存在顯 著差異，直接融合閃光燈特徵往往會引入光照相關的偽影，影響輸出品質。本論文提 出 SGAFNet，一個具備可信度感知能力的閃光燈引導去噪網路，透過選擇性地傳遞 可靠的結構線索，同時抑制兩張影像不一致的區域，以解決上述問題。本方法利用基 於梯度的可信度估計機制，對跨模態特徵融合進行空間控制，並結合注意力機制進行 特徵重校正，以及一個輔助損失函數來強化高可信度區域的結構一致性。在 FAID 與 MID 資料集上，於多個高斯噪聲等級下的實驗結果顯示，SGAFNet 在 PSNR 與 SSIM 兩項指標上均穩定優於所有比較方法，在高噪聲等級下的優勢尤為明顯。上述 結果驗證了明確建模引導可信度對於穩健的閃光燈引導影像去噪是一種有效的策略。;Flash-guided image denoising aims to recover clean ambient images by leveraging structural information from a paired flash image. However， directly fusing flash features into the denoising process often introduces illumination-induced artifacts， as the two images differ significantly in photometric appearance， flash shadows， and reflection patterns. In this thesis， we propose SGAFNet， a confidence-aware flashguided denoising network that addresses this challenge by selectively transferring reliable structural cues while suppressing regions where flash and ambient images disagree. Our method incorporates a gradient-based confidence estimation mechanism to spatially control cross-modal feature fusion， combined with attention-based feature recalibration and an auxiliary loss that enforces structural consistency in high-confidence regions. Additionally， we design a multi-scale progressive restoration module to better preserve high-frequency details during the denoising process. Experimental results on the FAID and MID datasets under multiple Gaussian noise levels demonstrate that SGAFNet consistently outperforms the compared baseline methods in both PSNR and SSIM， with especially clear advantages at higher noise levels. These results confirm that explicitly modeling guidance reliability is an effective strategy for robust flash-guided image denoising， offering a practical solution for low-light photography and computational imaging tasks.
