在現實生活中，拍攝影像常受到各種劣化因素的影響，進而導致成像品質下降。深度學習影像修復技術旨在透過神經網路模型，將低品質的劣化影像恢復為高品質的清晰影像。本博士論文聚焦於高效能深度學習影像修復之研究，涵蓋影像去模糊與惡劣天氣影像修復兩大領域。本文將針對高效能網路架構設計、影像修復資料集之擴充和領域自適應等議題進行深入探討。
由於影像修復模型常部署於邊緣裝置，如手機、相機與監視攝影機，如何設計兼具高成效與低資源消耗的修復模型，已成為本領域中的重要研究課題。在本論文的第一項研究中，我們觀察到雖然傳統 Transformer 架構具備強大的表現能力，但其高計算複雜度與對大規模訓練資料的依賴，限制了其在實際應用中的可行性。因此，本文提出 Stripformer，一種專為影像去模糊任務設計的高效能 Transformer 架構，Stripformer 採用結合水平方向與垂直方向自注意力機制的設計，能有效捕捉不同尺度與方向的模糊特徵；即便在不依賴大量訓練資料與不增加過多計算成本的情況下，仍能達成優異的去模糊效果，展現其於實務應用中的潛力。
除了僅處理單一劣化因素的影像去模糊任務外，多合一影像修復亦是近年來受到廣泛關注的研究方向，旨在以單一模型同時處理多種影像修復任務。鑑於現實生活中常見的惡劣天氣情況（如降雨、降雪與霧氣）之間具有高度相關性，且常同時發生，本論文的第二項研究聚焦於設計一個針對多種天氣劣化的多合一影像修復模型。為了有效修復受惡劣天氣影響的影像，我們提出三重注意力網路（Triplet Attention Network），以充分利用三種天氣條件之共通特徵，達成高效能的多任務影像修復任務。
另一方面，本文觀察到現有研究大多聚焦於模型架構的改進，以提升影像修復任務之性能。然而，除了設計更優化的網路架構外，訓練資料品質亦是影響模型表現的關鍵因素之一。基於此觀察，本文的第三項研究進一步探討如何在不增加運算複雜度的前提下，透過提升訓練資料品質以強化模型效能。為此，本文提出 ID-Blau，一種基於隱式擴散模糊影像生成模型，ID-Blau 能夠合成具備真實世界模糊特性的影像資料，並用以輔助模型訓練。實驗結果顯示，透過 ID-Blau 所生成的訓練資料，可顯著提升多種影像去模糊模型的修復品質，驗證其在資料層面的有效性與通用性。
最後，在本文的第四項研究中，本文進一步探討影像去霧任務中資料不足的挑戰。現有真實世界影像去霧資料集數量有限，導致訓練出的模型在真實場景下表現不佳，嚴重限制其應用效能。為解決此問題，本文提出一個霧氣轉換模型:PHATNet，用以解決霧氣修復模型的領域自適應問題。由於現實世界中的霧氣型態具有高度多樣性與不可控性，PHATNet 能夠從測試影像中提取霧氣特徵，並轉移至訓練資料中的乾淨影像上，合成具有目標測試環境霧氣特徵的影像資料。藉由此方法，PHATNet 可用於產生具針對性的霧氣資料集，進一步用於微調現有霧氣修復模型。實驗結果顯示，此領域自適應可顯著提升模型在未知測試場景下的去霧表現，展現其強大的跨域適應能力與實用性。

Images captured in real-world scenarios often suffer from various types of degradation, leading to a reduction in image quality. Image restoration aims to recover high-quality, clean images from degraded inputs. This dissertation focuses on efficient deep learning-based image restoration, with an emphasis on two major tasks: image deblurring and image deweathering. In particular, we focus on two key aspects for addressing image restoration: attention networks for image restoration and physics-guided data augmentation and adaptation for image restoration.
Since image restoration models are commonly deployed on edge devices such as smartphones, cameras, and surveillance systems, designing efficient and effective models has become a crucial research challenge in this field. Therefore, we propose two efficient attention modules for image deblurring and deweathering. In our first work, inspired by the rapid development of Transformer, we observe that although standard Transformers exhibit powerful modeling capabilities, they often suffer from high computational complexity and heavily rely on large-scale training data. Therefore, we propose Stripformer, an token-efficient and parameter-efficient Transformer designed for image deblurring. Stripformer incorporates the inductive bias of image deblurring to efficiently captures blur patterns of varying orientations and magnitudes, achieveing strong deblurring performance without relying on large-scale datasets.
Additionally, beyond addressing a single degradation type, several studies have explored unified approaches for handling multiple degradations. Among them, adverse weather conditions, such as rain, snow, and haze, are closely related and often co-occur, making them suitable for joint restoration. Therefore, in our second work, we propose a Triplet Attention Network (TANet) that leverages shared characteristics across different weather-related degradations to efficiently perform all-in-one adverse weather image restoration.
Moreover, we observe that most existing works primarily focus on architectural improvements to enhance restoration performance. However, beyond model design, the quality of training data also plays a crucial role in determining restoration effectiveness. Motivated by this insight, we investigate how to improve restoration results by enhancing data quality without increasing the computational burden of restoration models. Therefore, in our third work, we propose ID-Blau, a blur augmentation method based on an implicit diffusion-based reblurring network. ID-Blau generates blurred images that simulate realistic and diverse blur patterns, which are then used to augment the training data of deblurring models. This strategy significantly improves restoration performance while maintaining computational efficiency.
Lastly, in our last work, we further investigate the challenge of limited data in image dehazing tasks. Real-world image dehazing datasets are scarce, which limits the generalization ability of dehazing models to unseen target domains. To address this issue, we propose PHATNet, a haze transfer network that transfers haze patterns from unseen target domains to clean images in the source domain. This approach enables us to generate domain-adaptive fine-tuning sets, which can be used to adapt dehazing models at test time. As a result, the proposed method significantly improves dehazing performance in target domains.