ShabbyPages

ShabbyPages是由Sparkfish LLC和范德比尔特大学合作创建的一个用于文档去噪和二值化的大型数据集。该数据集包含6202张来自多种语言和格式的‘生数字’图像及其合成噪声版本，旨在模拟打印、传真等物理过程后的文档状态。数据集通过Augraphy工具进行增强，以增加其复杂性和多样性，适用于训练和评估现代数据驱动的机器学习模型。ShabbyPages不仅支持文档处理技术的发展，还为相关领域的研究提供了丰富的资源，特别是在光学字符识别和布局解析等下游任务中。ShabbyPages is a large-scale dataset for document denoising and binarization, developed in collaboration between Sparkfish LLC and Vanderbilt University. It contains 6202 'raw digital' images and their synthetic noisy counterparts across diverse languages and formats, designed to simulate document states resulting from physical processes such as printing and faxing. The dataset is augmented using the Augraphy tool to enhance its complexity and diversity, making it suitable for training and evaluating modern data-driven machine learning models. ShabbyPages not only supports the advancement of document processing technologies, but also provides abundant resources for research in related fields, particularly in downstream tasks such as optical character recognition (OCR) and layout parsing.