A semantic and attribute-based dataset for image super resolution

Despite the remarkable advancements propelled by deep learning in Image Super Resolution, the enigmatic essence of deep neural networks introduces ongoing challenges, particularly emphasizing the necessity for interpretability in the elucidation of how image features are transmuted into enhanced high-resolution outputs. With the strategic importance of training datasets in view, we introduce a rigorously curated UHD image dataset endowed with distinct attributes. This dataset, systematically categorized and annotated, includes a diverse assortment of semantic textures such as windowpanes, trees, skyscrapers, and architectural elements. Its design is intended to streamline the training and testing of a variety of SISR models, accommodating numerous resolution scales and image formats predicated on high-quality original images. It is developed with the goal of improving the identification of standard mapping functions within SISR, thus facilitating the generation of more productive and versatile models.