超解析度成像法的目的是提升影像的空間解析度，恢復影像次像素中遺失的細節。傳統提升影像解析度的方法多使用線性或非線性內插，運行快速、耗費運算資源低，但在影像成果較模糊，高頻細節的復原並不理想。隨著類神經網路的發展，針對影像處理的卷積神經網路大量運用於超解析度成像法。許多卷積神經網路的超解析度成像法為監督式學習方法，需要大量成對的外部影像集將模型訓練至可提升特定特徵目標如人臉、風景或建物的影像解析度，因此訓練過程運算需求高且耗時長，訓練成果與外部影像集的豐富程度和訓練迭代次數相關。相對而言，非監督超解析度成像法不需要耗費高成本獲得成對的外部影像集來學習影像從低解析度到高解析度的映射，大幅減少計算資源，使非監督學習對未知的新數據有更好的適應性，提高實際應用場景中的可用性和應用性。 本論文針對三種非監督學習方法Zero-shot Super-resolution (ZSSR)、Deep Image Prior (DIP) 及Degradation-aware Super-resolution (DASR)，比較其提升Sentinel-2及SPOT衛星影像之空間解析度的效能。實驗結果顯示，非監督學習方法相較於雙三次內插法在峰值訊噪比(Peak Signal-to-Noise Ratio, PSNR)、結構相似性(Structural Similarity, SSIM)及特徵相似性(Feature Similarity, FSIM)有所提升，與耗費訓練資源的監督式學習方法相比仍展現優勢。成果也顯示三種非監督學習方法中，Degradation-aware Super-resolution在光學衛星影像中表現最佳。 ;The purpose of super-resolution imaging is to improve the spatial resolution of images, restore lost details in the sub-pixel levels. The traditional methods based on linear or nonlinear interpolations are fast and consume fewer computing resources. But the results are usually blur and not suitable for high-frequency details recovery. With the development of artificial intelligent, the convolutional neural networks (CNN) are widely adopted for image super-resolution. Most of CNN-based super-resolution algorithms are supervised learning. They need huge amount of external training image pairs to train the model and improve the image resolution for specific feature such as faces, scenery or buildings. Therefore, they all require high computational power and long training time. The richness of the external image set and the number of training iterations are related to the effectiveness. On the other hand, unsupervised super-resolution algorithms aim to learn the relationship from low-resolution to high-resolution images only from the test image itself. Unsupervised methods do not need to collect paired training data, and can reduce computing resources significantly. This also makes a flexible adaption on new and unseen data, increased accessibility and applicability in real-world scenarios. In this study, three different unsupervised super-resolution methods, Zero-shot Super-resolution (ZSSR), Deep Image Prior (DIP) and Degradation-aware Super-resolution (DASR), are implemented and applied to improve the spatial resolution of Sentinel-2 and SPOT images. The experimental results show that unsupervised methods can improve Peak Signal-to-noise Ratio (PSNR), Structural Similarity (SSIM) and Feature Similarity (FSIM) compared with bicubic interpolation, also competitively to supervised methods that consume expansive training resources. The results also show that Degradation-aware Super-resolution performs the best among these three unsupervised methods.
