DRID-20

DRID-20是由悉尼科技大学全球大数据技术中心开发的一个图像数据集，包含20个类别，总计20000张图像。该数据集通过利用Google Books Ngram Corpus进行查询扩展，筛选出视觉上显著且相关的图像，以减少数据集偏差。创建过程中，每个查询扩展被视为一个‘包’，检索到的图像作为‘实例’，通过多实例学习方法进行图像选择和噪声过滤。DRID-20旨在提高图像分类和跨数据集泛化的能力，适用于多实例学习、迁移学习和图像数据集构建等研究领域。

DRID-20 is an image dataset developed by the Global Big Data Technology Center of the University of Technology Sydney. It contains 20 categories and a total of 20,000 images. To mitigate dataset bias, this dataset leverages the Google Books Ngram Corpus for query expansion, and screens visually salient and relevant images. During the creation process, each query expansion is treated as a 'bag', while the retrieved images are regarded as 'instances', and image selection and noise filtering are conducted through multiple instance learning methods. DRID-20 is designed to enhance the performance of image classification and cross-dataset generalization, and is applicable to research fields including multiple instance learning, transfer learning, and image dataset construction.