葉家宏Yeh, Chia-Hung王宇捷Wang, Yu-Chieh2023-12-082023-01-052023-12-082023https://etds.lib.ntnu.edu.tw/thesis/detail/6692038bb3bb12608e9631c100e43b40/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/120326在影像運動去模糊中，多階段架構已被廣為使用且獲得卓越的效能。而過去傳統方法通常透過提取模糊輸入影像的空間細節來修復退化的影像，但由於輸入模糊圖像無法提供準確的高頻細節，因此降低了整體去模糊演算法的性能。為了解決這個問題，本文提出一種新的雙階段架構，該架構可以通過提取模糊圖像的高頻細節資訊來重建詳細的紋理。它使用了一種有監督的引導機制來提供精確的空間細節來重新校準多尺度特徵。此外，該方法還設計一個基於注意力的特徵聚合器，可以自適應地融合來自不同階段的有影響的特徵，以抑制傳遞到下一個階段的冗餘資訊。本文通過實驗證明，所提出的法在 GoPro 和 HIDE 基準數據集上的去模糊性能和計算複雜度方面都優於其他現有方法。Multi-stage image processing architectures have been widely used for deblurring and have shown good results. Traditional methods try to restore the details of the blurred image by using the information in the blurred image itself, but this can be problematic because the blurred image may not contain enough high-frequency detail, leading to a loss of quality in the deblurred image. To solve this problem, we propose a dual-stage network that is able to retrieve high-frequency detail from the blurred image and use it to restore the image's texture in more detail. We also introduce a mechanism that uses known image content to better calibrate contextual information, and an attention-based feature aggregator that adaptively combines influential features from different stages to eliminate unnecessary information and improve the efficiency of the multi-stage architecture. Our experiments on GoPro and HIDE datasets show that our network performs better and is more efficient than existing methods.影像去模糊非均勻運動模糊去除圖像恢復高頻資訊Image DeblurringNon-uniform motion deblurringImage RestorationHigh-frequency informationetd