Phips/BHI

BHI SISR数据集用于训练单图像超分辨率模型，包含390,035张512x512像素的webp格式图像。该数据集通过BHI过滤方法优化，确保训练质量。数据集的优势在于其大量的标准化训练图像，适用于多种超分辨率模型的训练。此外，还提供了对应的低分辨率数据集和训练模型。

The BHI SISR Dataset is designed for training single image super-resolution models and consists of 390,035 images, all 512x512 pixels in size and stored in the webp format. The dataset is the result of applying the BHI filtering method to various high-quality datasets, which removes the worst quality tiles to improve training metrics. Its advantage lies in the large quantity of normalized training tiles, ensuring a sufficient distribution of degradation strengths when creating corresponding low-resolution datasets for paired training. The dataset is composed of multiple source datasets, processed through tiling and filtering, with files named {dataset_name}_{index}.webp. Additionally, corresponding low-resolution datasets and models trained on this dataset are provided.