Real-CE
收藏资源简介:
Real-CE数据集是由香港理工大学和OPPO研究院联合创建的,专注于中英文场景文本图像超分辨率的实际应用。该数据集包含1,935对训练图像和783对测试图像,总计33,789条文本行,其中24,666条为中文,9,123条为英文。数据集通过多种场景和资源收集,确保了文本内容、呈现和光照条件的多样性。创建过程中,使用了不同焦距的相机模块来捕捉图像,并通过图像注册方法进行对齐。Real-CE数据集旨在解决现有数据集在处理结构复杂的中文文本时的不足,通过提供详细的标注信息,如检测框和文本转录,来辅助训练和评估。此外,数据集还设计了适应不同文本行大小的评估流程,以保持超分辨率文本图像的视觉质量。
Real-CE Dataset was jointly created by The Hong Kong Polytechnic University and OPPO Research Institute, focusing on practical applications of super-resolution for scene text images in both Chinese and English scenarios. This dataset contains 1,935 pairs of training images and 783 pairs of test images, with a total of 33,789 text lines, among which 24,666 are Chinese and 9,123 are English. The dataset is collected from diverse scenes and resources to ensure the diversity of text content, presentation styles and lighting conditions. During the creation process, camera modules with different focal lengths were used to capture images, and alignment was performed via image registration methods. The Real-CE Dataset aims to address the shortcomings of existing datasets when handling structurally complex Chinese text, and assists model training and evaluation by providing detailed annotation information such as detection bounding boxes and text transcriptions. Furthermore, the dataset designs an evaluation pipeline adapted to different text line sizes to maintain the visual quality of super-resolved text images.

- 1A Benchmark for Chinese-English Scene Text Image Super-resolution香港理工大学; OPPO研究院 · 2023年



